text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
null | null |
# Introduction {#sec:intro}
Direction of arrival (DoA) estimation has been the subject of much recent research, particularly in the context of channel characterization for wireless communications . Conventional receiver architectures [for]{style="color: black"} DoA estimation leverage antenna array-based solutions with a dedicated signal processing layer, such as MUSIC , ESPRIT , and SAGE . [Antenna arrays typically have elements separated by \(\lambda/2\) where \(\lambda\) is free-space wavelength.]{style="color: black"} As a result, array-based topologies require that the received signal is collected through each antenna element within the array. Such an approach can significantly increase the hardware complexity due to an excessive number of data acquisition channels, particularly when the electrical size of the aperture is large and the operating frequency is increased. Recently, the concept of compressive sensing has received significant [attraction]{style="color: black"} as an enabling technology [for DoA estimation. In, the compressive DoA estimation study was facilitated using a phased array aperture. The phased array aperture requires that each array element has a dedicated phase shifting circuit. As an alternative approach, single-pixel, wave-chaotic metasurface antennas have recently been shown to offer significant potential for compressive sensing, particularly in millimeter-wave (mmW) imaging . A significant advantage of the single-pixel coded-aperture concept is that these apertures do not require a dedicated feeding network to control the excitation phase of each array antenna element]{style="color: black"}. [These metasurface apertures are called wave-chaotic since they can radiate spatially-varying, quasi-random radiation patterns over time. These radiation patterns are called spatio-temporally incoherent because of the low level of correlation that exists between the radiation patterns synthesized by the metasurface aperture.]{style="color: black"} The underlying principle behind this concept is that, the scene information can be encoded onto spatio-temporally incoherent wave-chaotic modes radiated by single-pixel compressive coded-aperture antennas. [It has been proven that such an approach can substantially simplify the physical layer architecture.]{style="color: black"} Leveraging the fundamentals of the compressive sensing concept, the authors previously developed a compressive DoA estimation technique using a single-pixel coded-aperture and a [sparse array-based]{style="color: black"} receiver topology . The single-pixel technique can successfully retrieve the DoA information of sources arbitrarily defined in the far-field of the aperture using only a single channel to acquire the data. Despite the success of these proof-of-concept studies and encouraging results obtained in these works, compressive DoA estimation concept currently suffers from two limitations: First, a priori knowledge of the fields radiated by the compressive antenna must be known. This suggests that a characterization step is needed to obtain the transfer function of the antenna. Second, the compressed signal received at the antenna channel needs to be interacted with the transfer function of the antenna to recover the DoA pattern. This suggests that a reconstruction step is needed. Both these steps are computationally expensive and pose a significant challenge for practical applications. The main motivation of this work is to develop a single-pixel DoA estimation technique leveraging a deep-learning layer to directly estimate the DoA information from the compressed channel measurements, eliminating the need for the reconstruction step altogether. To this end, we develop a Graph Attention Network (GAT) approach embedded within a coded-aperture-based, single-pixel radar framework for DoA estimation. The contributions of the presented study are summarized as follows:
1. We propose a compressive, single-pixel framework that can significantly simplify the receiver physical layer architecture for DoA estimation.
2. We present the first deep learning-based approach as applied to [coded-aperture-based]{style="color: black"} compressed channel measurements for DoA estimation. The proposed method is applied directly to the raw data measured at the compressed channel of the single-pixel receiver.
3. We analyze the performance of the GAT integrated single-pixel compressive DoA estimation technique under various signal-to-noise ratio (SNR) level conditions to present the potential of this technique, [as a proof-of-concept]{style="color: black"}, for channel characterization.
The outline of this paper is as follows: In Section [2](#sec:preliminaries){reference-type="ref" reference="sec:preliminaries"}, we explain GAT and the concept of compressive sensing as applied to DoA estimation. In Section [3](#sec:method){reference-type="ref" reference="sec:method"}, we present the deep learning-based single-pixel DoA estimation scheme integrated with GATs as an enabling learning approach. Finally, in Section [5](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}, we provide the concluding remarks.
# Preliminaries {#sec:preliminaries}
## Graph Attention Networks {#sec:gat}
[Conventional deep learning methods such as CNN can provide accurate results on challenging tasks.]{style="color: black"} However, they suffer from data without grid-like structure . As a state-of-the-art method, [GNNs]{acronym-label="GNN" acronym-form="plural+short"} have been recently proposed to deal with data without grid-like structure . By using graph structure and node features, [GNNs]{acronym-label="GNN" acronym-form="plural+short"} can learn the representations of nodes and the graph. Furthermore, the composition of [GNN]{acronym-label="GNN" acronym-form="singular+short"} with attention mechanism creates a more robust method, namely [GATs]{acronym-label="GAT" acronym-form="plural+short"}, with inductive learning capability . Therefore, it is expected that [GATs]{acronym-label="GAT" acronym-form="plural+short"} can be utilized for DoA problems due to the random position of the source and random propagation environment. The fundamental part of a [GAT]{acronym-label="GAT" acronym-form="singular+short"} is [GAL]{acronym-label="GAL" acronym-form="singular+short"} with \(K\) input nodes. The input nodes are defined as follows: [\[\begin{aligned}
\mathbf{\gamma}=\left\{\vec{\gamma}_{1}, \vec{\gamma}_{2}, \ldots, \vec{\gamma}_{K}\right\}, \, \vec{\gamma}_{i} \in \mathbb{R}^{F}, \end{aligned}\]]{style="color: black"} where \(F\) denotes the number of features in each node. At the output of [GAL]{acronym-label="GAL" acronym-form="singular+short"}, the cardinality of the output set might differ from the input set. Therefore, let us denote the number of features of nodes as \(F^{\prime}\) without loss of generality. In each [GAL]{acronym-label="GAL" acronym-form="singular+short"}, a linear operation that is defined by a weight matrix, [\(\mathbf{W} \in \mathbb{R}^{F^{\prime}\times F}\)]{style="color: black"}, transforms the input attributes of nodes to higher-level attributes. Following the linear transformation, self-attention on nodes is performed by the self-attention mechanism, \(a: \mathbb{R}^{F^{\prime}} \times \mathbb{R}^{F^{\prime}} \rightarrow \mathbb{R}\). The attention mechanism computes attention coefficients [according to]{style="color: black"}: \[\begin{aligned}
c_{i j}=a\left(\mathbf{W} \vec{\gamma}_{i}, \mathbf{W} \vec{\gamma}_{j}\right), \end{aligned}\] where the neighborhood between the \(i\)-th and \(j\)-th nodes is shown by \(c_{i j}\). In other words, the attention coefficient shows how much of an effect the attributes of the \(j\)-th node have on the \(i\)-th node. By using the softmax function, we normalize the coefficients as follows: \[\begin{aligned}
\alpha_{i j}=\operatorname{softmax}_{j}\left(c_{i j}\right)=\frac{\exp \left(c_{i j}\right)}{\sum_{k \in \mathcal{N}_{i}} \exp \left(c_{i k}\right)}, \end{aligned}\] where \(\mathcal{N}_{i}\) denoted neighborhood of \(i\)-th node, which is defined by a binary adjacency matrix \(\mathbf{A} \in\{0,1\}^{K \times K}\). Then, the convolution operation is performed over [the]{style="color: black"} graph[, and modeled as:]{style="color: black"} \[\begin{aligned}
\mathbf{Z}=\alpha \mathbf{X} \mathbf{W}+\mathbf{b}, \end{aligned}\] where \(\mathbf{X} \in \mathbb{R}^{K \times F^{\prime}}\) and \(\mathbf{b}\) are the node attributes matrix and the trainable bias vector, respectively. Finally, the graph is reduced to a single vector by global pooling layer . It enables generalization of the network and decreases the number of representations. Thus, global pooling layer can avoid the network from overfitting. For more information on [GAT]{acronym-label="GAT" acronym-form="singular+short"} and its previous application on channel estimation, we refer the readers to .
## Compressive Sensing and DoA Estimation {#sec:compressing_sensing}
Compressive sensing is a computational technique that can be used to solve a sparse problem from a reduced number of measurements . In , we demonstrate the compressive aperture concept facilitated for the DoA estimation problem. [In the presented concept, the coded-aperture works as a receiver whereas the far-field source works as a transmitter. It should be noted that, if implemented in a wireless communications scenario, such as channel characterization, the proposed technique can be considered suitable for implementation at the network side to identify the DoA information of far-field sources that are incident on the coded-aperture.]{style="color: black"} The size of the synthesized aperture is \(D=25\) cm and the operating frequency is chosen to be 28 GHz. The compressive aperture depicted in is a coded-aperture, suggesting that it radiates spatio-temporally incoherent radiation patterns by means of reconfiguring the aperture . In this concept, each aperture reconfiguration state can be considered a *mask* with each mask radiating a different radiation pattern. [A careful investigation of the wave-chaotic radiation patterns depicted in reveals that the radiation pattern for each mask can be considered [as]{style="color: black"} a collection of sidelobes, rather than a well-defined main-lobe with a fixed 3dB beamwidth.]{style="color: black"} For the compressive DoA system depicted in , we use a total number of 2000 masks. [It should be noted that the dynamic modulation of the coded-aperture can be realized using transistors. These elements exhibit an extremely fast switching response, typically on the order of a few nanoseconds. As a result, sweeping through 2000 masks can be achieved over a microsecond time-frame. This is smaller than the typical coherence time for 5G channels. The selection of the masks is done on a random basis.]{style="color: black"} As shown in , the aperture has a single channel for data acquisition. A significant advantage of the physical layer compression depicted in can be appreciated when considering the same size aperture synthesized using an array topology. At the conventional \(\lambda/2\) limit, 0.54 cm, the aperture in would require 2209 elements within the array. Considering the compressive DoA estimation scenario presented in , the measured signal at the compressed channel can be represented as follows: \[\textcolor{black}{ \mathbf{s}_{M\times1}=\mathbf{E}_{M\times N}\mathbf{P}_{N\times1}+\mathbf{n}_{M\times1}}. \label{Equation_6}\] Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}) is known as the forward-model. In Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}), **s** represents the acquired data at the compressed channel, **E** is the radiated field (or the transfer function) of the coded-aperture, **P** is the projection of the far-field source (or sources) on the aperture of the antenna and **n** is the noise, which controls the SNR level in the acquired data. It should be mentioned that the bold font here denotes the vector-matrix notation. In Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}), \(M\) and \(N\) denote the number of masks and [number of pixels in the projection of the far-field source on the antenna aperture]{style="color: black"}, respectively. [In this context, a further insight into the **E** and **P** is needed to appreciate the physical layer compression concept. The transfer function of the antenna across a plane in front of its aperture can be calculated as :]{style="color: black"} \[\begin{aligned}
\textcolor{black}{ \mathbf{E(\mathbf{r})}=i\frac{\omega \mu_{0}}{4\pi}{\sum_{n}(\mathbf{m}_{n}\times\mathbf{r})}\Bigg(\frac{ik}{R_{n}}-\frac{1}{R_{n}^2}\Bigg)e^{jkR_{n}}}, \label{Equation_E} \end{aligned}\] [where \(\bf{r}\) denotes the coordinates across the plane defined in front of the coded-aperture, \(R_{n}=|\bf{r}-\bf{\rho}_{n}|\), and \(\bf{\rho}_{n}\) and \(\bf{m}_{n}\) denote the coordinates and the magnetic dipole of the \(n^{th}\) metamaterial element in the metasurface layer. Similarly, the projection of the far-field source across the same plane in front of the coded-aperture is calculated as follows :]{style="color: black"} \[\begin{aligned}
\textcolor{black}{ \mathbf{P(\mathbf{r})}=e^{-jk_{0}(y\sin{\theta}\cos{\varphi})(z\sin{\theta}\sin{\varphi})}}, \label{Equation_P} \end{aligned}\] [In Eq. ([\[Equation_P\]](#Equation_P){reference-type="ref" reference="Equation_P"}), \(\theta\) and \(\varphi\) are the incident angles of the far-field source while \(k_0\) is the wavenumber. In order to visualize the relationship between **s**, **E** and **P**, in Fig. [\[fig:s_E\_P\]](#fig:s_E_P){reference-type="ref" reference="fig:s_E_P"}, we provide a depiction of the mathematical model for the compressive DoA estimation concept. As depicted in Fig. [\[fig:s_E\_P\]](#fig:s_E_P){reference-type="ref" reference="fig:s_E_P"}, the mapping of \(\bf{P}\) onto \(\bf{s}\) is achieved through the transfer function of the coded-aperture, \(\bf{E}\).]{style="color: black"} From Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}), by performing a phase compensation on the transfer function and applying the compensated transfer function to the compressed channel data, an estimate of the field projection pattern, \(\textbf{P}_{est}\), can be retrieved. To achieve this, several computational techniques, such as the least-squares method minimizing the objective function \(||\mathbf{E}^\dagger\mathbf{s}-\mathbf{P}||_2^2\), where \(.^\dagger\) denotes the phase conjugation operator and \(||.||_2\) denotes the Euclidean operator, can be used. Analyzing Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}), it can be seen that there are two [fundamental]{style="color: black"} challenges with the compressive DoA estimation technique. First, it is evident that the signal captured at the coded-aperture antenna channel is correlated to the far-field source through the transfer function of the antenna. As a result, achieving the DoA estimation using the compressive coded-aperture technique requires a [priori]{style="color: black"} knowledge of the antenna transfer function. This step makes it necessary to measure the radiated fields from the coded-aperture as part of a calibration step, requiring a hardware-intense, highly complex characterization process . Moreover, in characterizing the transfer function of the antenna, an excellent phase accuracy is needed. [This is because]{style="color: black"} any errors introduced to the transfer function of the antenna during the characterization process can significantly distort the forward-model in Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}) . Second, the DoA estimation requires that Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}) [is]{style="color: black"} solved to recover \(\textbf{P}_{est}\). This reconstruction step can take a considerable amount of time, which is of particular concern for applications where the DoA estimation needs to be done in real-time. Alternatively, retrieving the DoA information directly from the compressed channel measurements can eliminate the [need to know the antenna transfer function]{style="color: black"} and the necessity for the reconstruction step.
# Single-Pixel 2-D DoA with GATs {#sec:method}
To address the challenges of the compressive DoA estimation concept, it is desirable that an alternative technique is developed that can work with the channel data without the need to process the forward-model of Eq. ([\[Equation_6\]](#Equation_6){reference-type="ref" reference="Equation_6"}). Such a technique can eliminate the need to consider the DoA estimation problem as a reconstruction problem. Therefore, the ultimate goal of this work is to develop a [GAT-based]{style="color: black"} deep learning solution that can directly learn from the compressed channel data. GAT has been recently proposed for learning over graph-structural data by exploiting the self-attention mechanism. In this study, GATs are utilized to deduce the relation between compressed channel data and DoA of the received signal. First, we converted the received compressive signal into graph structural data given with \(\mathbf{X}\) by mapping the imaginary and real parts of the signal into the attributes of graph nodes as follows: [\[\begin{aligned}
\mathbf{X} =[\operatorname{Re}\{\mathbf{s}^\top\}; \operatorname{Im}\{\mathbf{s}^\top\}], \end{aligned}\] where, \((\cdot)^\top\) denotes matrix transpose.]{style="color: black"} Also, we defined adjacency relation between both nodes with an adjacency matrix, \(\mathbf{A}\), which is a \(2\times2\) minor diagonal matrix (i.e., \(K=2\)). \(\mathbf{A}\) defines a graph without any self-linked edge. In other words, only a simple graph is defined. This simple graph allows a relationship to be held in terms of phase information over the relation between real and imaginary attributes of the nodes. Moreover, it allows to [weigh]{style="color: black"} this information with an attention mechanism. The label input is also defined as: [\(\mathbf{y} = [\mathbf{\theta};\, \mathbf{\varphi}]\)]{style="color: black"}. In this study, \(4\) consecutive [GALs]{acronym-label="GAL" acronym-form="plural+short"} are employed to learn features over the graph. In the sequel, we used a graph attention pooling to generalize the network and avoid it from overfitting. The proposed model also utilizes \(4\) dense layers for the regression of the compressed signals over the angle pairs. [The proposed [GAT]{acronym-label="GAT" acronym-form="singular+short"} architecture for 2-D DoA is illustrated in .]{style="color: black"} [For nonlinearity,]{style="color: black"} ReLU is chosen as [an]{style="color: black"} activation function in each layer. [MAE]{acronym-label="MAE" acronym-form="singular+short"} is used as a loss function in this network since it provides a suitable metric for use in DoA estimation problems by its nature. As RMSprop has been proposed for fast and robust optimization , it was preferred for the optimization method of the model in this study. The learning rate of [the]{style="color: black"} optimizer has been set to \(5\times10^{\minus6}\). It is worth noting that performance evaluation of other optimization methods for the problem addressed by this study is an open issue for future studies. Applying the procedure mentioned below, we created the training and test datasets for varying numbers of antenna masks. Hence, it should be noted that the number of features of each node, \(F^{\prime}\), [becomes]{style="color: black"} \(M\). For the training dataset, compressed signals with both \(50\) dB SNR and 2-D DoA values (i.e., azimuth and elevation angles), which are uniformly distributed between \(\minus30\) and \(30\) with \(1\) steps, were prepared using varying numbers of antenna masks. [To ensure that the number of received signal samples for each azimuth and elevation angle pair is sufficient for learning, we chose the number of sample compressed signals as \(50000\). In other words, for training, the number of signal samples for each angle pair is above 10]{style="color: black"}. The dataset is publicly available in . The network is trained with the data detailed above through \(500\) epochs unless early stopping is activated. Non-decreasing loss for \(20\) epochs evokes early stopping in this configuration. Thus, early stopping, besides \(L_2\)-regularization, prevents the network from overfitting.
# Numerical Results {#sec:result}
The performance evaluation of the trained model is investigated for [a]{style="color: black"} various number of antenna masks. For performance evaluation, a test dataset was created at varying SNR levels and number of antenna masks. The test dataset consists of \(5000\) single-pixel compressed signals received by metasurface aperture at each SNR level that is between \(\minus20\) and \(30\) dB. It should be noted that despite using signals at a high SNR level (i.e. \(50\) dB) during training, the test process has been carried out under more challenging conditions (i.e. lower SNR). As detailed in , test data was also transformed to graph representation by defining the nodes and adjacency relation. \(5000\) signals almost guarantee that one of each pair of angles is in the set. Thus, the results show an average performance for whole spatial scope. The initial results regarding the proposed proof-of-concept method demonstrate robust and high performance without any knowledge of the antenna transfer function and reconstruction step. The results given in show that even when the number of antenna masks is limited to \(100\), [GAT can estimate the elevation and azimuth DoAs of the source with a mean absolute error (MAE) of \(2.5\) and \(1.4\), respectively]{style="color: black"}. Increasing the number of antenna masks improves the estimation accuracy whereas the computational complexity also increases. [For example, GAT using more than \(500\) masks achieves less than \(2\) MAE at \(10\) dB SNR for both elevation and azimuth]{style="color: black"}. It is observed that the estimation performance cannot improve while SNR value is increasing beyond \(10\) dB. The reason for this phenomena is that the spatial scope is sampled with \(1\) step. Spatial sampling being a limiting factor can be clearly seen in . It is known that the smallest [MAE]{acronym-label="MAE" acronym-form="singular+short"} value for \(1\) interval is \(0.5\). shows that the results converge to the optimum value when more than \(10\) dB SNR is used and more than \(100\) masks are used. Oversampling might improve the estimation accuracy of [GAT]{acronym-label="GAT" acronym-form="singular+short"}, however, the training process would require more time and higher computational capacity due to increased data size. As seen in the results, the elevation estimation is slightly less accurate than the azimuth. The phenomena behind this situation can be explained by the [mask's]{style="color: black"} antenna pattern diversity in azimuth and elevation. It can be said that the radiation patterns of the designed metasurface aperture are more diverse in azimuth. In other words, as random measurement matrices improve sensing performance , it can be considered that antenna masks for compressive sensing include more randomness in azimuth. [Although the aperture size and shape are the same in the azimuth and elevation direction, the metamaterial elements in the metasurface aperture can have an asymmetric radiation pattern .]{style="color: black"} [Furthermore, the comparative results for azimuth and elevation estimations with \(1000\) masks are provided in . It is demonstrated that the GAT technique shows a higher performance than conventional DL methods at low SNR regimes because of its attention mechanism. Morevoer, the proposed method has lower training time-complexity. For example, while the duration per epoch for GAT is \(\tau\), AMPBML , DeepBeam , and CNN require \(1.8\tau\), \(19.6\tau\), and \(3.25\tau\), respectively. Here, \(\tau\) is the duration per epoch and depends on the computational capacity.]{style="color: black"}
# Concluding Remarks {#sec:conclusion}
We presented a single-pixel DoA estimation technique leveraging GATs to learn directly from the compressed channel measurements to retrieve the DoA information. It was shown that the developed technique can achieve DoA estimation from the compressed measurements without the necessity for a reconstruction step. As a particular example, we showed that, using \(M=2000\) aperture mask configurations, at an SNR level of \(10\) dB, the estimation error in the azimuth and elevation angles were below \(2\), confirming the potential of the presented technique for high-fidelity DoA estimation using a single-pixel receiver architecture and without the need for an additional reconstruction step. [It was presented that the proposed method improves the estimation performance in 2-D at low SNR regimes compared to conventional DL methods. As stated, the attention mechanism can leverage the node attributes. Therefore, it can be argued that GAT networks can perform satisfactorily under a scattering environment. This aspect will form the basis of our future studies.]{style="color: black"} | {'timestamp': '2021-11-03T01:03:39', 'yymm': '2109', 'arxiv_id': '2109.05466', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05466'} |
# Introduction {#Introduction}
The variational reformulation of Griffith's theory proposed by Francfort and Marigo and followed by Bourdin et al.'s phase-field regularization offered a robust mathematical and computational framework for the modeling of quasi-static brittle fracture. Based on a smooth representation of the discontinuities through a damage field, variational phase-field models compute the nucleation and propagation of complex crack paths through the minimization of an energy functional. As opposed to discrete approaches dealing explicitly with cracks like Linear Elastic Fracture Mechanics (LEFM) or Cohesive Zone Models (CZMs), the phase-field method for fracture can model crack propagation without the need for remeshing, a priori knowledge of the crack path or ad hoc criteria . Owing to its elegant simplicity and its compatibility with classical finite element methods, the phase-field method has been applied and extended to hydraulic fracture , ductile fracture , stress corrosion cracking and brittle fracture in heterogeneous structures . A well-known downside of phase-field models is their high computational cost. The first explanation for this situation is the necessity of a very fine mesh for the finite element method to accurately capture the strong damage gradient in the fracture zone, yielding large linear systems to be solved. Adaptive mesh techniques and multiscale methods have been developed and appear to successfully reduce the number of degrees of freedom to a minimum. The second explanation for the high computational cost is the non-linear and non-convex nature of the underlying energy functional of phase-field models . For non-convex problems, the conventional Newton's method is inadequate since the Hessian can become indefinite and hinder the convergence to a critical point. The first method proposed in the literature to solve phase-field models was the alternating minimization algorithm . Stemming from the observation that the functional is convex in each variable when the other field is fixed, the alternating minimization scheme alternatively solves the kinematic and damage sub-problems until convergence. Widely adopted by the community , the algorithm is robust, but exhibits a slow convergence rate due to the decoupling of the sub-problems . Significant gains in efficiency were obtained using a combination of preconditioned alternating minimization scheme and classical Newton's method , or a semi-implicit form of the staggered scheme . Nevertheless, such methods require either the tuning of or a time step convergence study, implying additional computation, while still suffering from the decoupling between the displacement and damage fields. . The quasi-monolithic scheme was also reported efficient and robust in , but the relaxation of the applied forces and a delayed crack growth was observed. Furthermore, the delayed and relaxed evolution seems to be amplified under unstable crack growth, where the damage field evolves rapidly. This suggests that the quasi-monolithic scheme generates a numerical viscosity, requiring smaller time steps and, therefore, an increased computational cost to accurately capture rapidly growing cracks. To the best of our knowledge, the efficiency of the alternating minimization algorithm and the quasi-monolithic scheme has never been compared. As reported by Wick , obtaining a monolithic scheme for phase-field models for fracture is a difficult task. Numerous monolithic algorithms have been proposed, such as a Newton method using a modified line-search algorithm , an error-oriented Newton's method , a modified Newton method with Jacobian modification , . Probably due to their complexity, variable robustness, or problem-dependent efficiency, these monolithic methods have yet to impose themselves, with the alternating minimization algorithm and quasi-monolithic scheme still being vastly more represented in the literature. More recently, Wambacq et al. proposed a monolithic solver relying on a modified Newton method with inertia correction inspired by . Although the modified Newton method showed interesting performance, their work focused on the efficiency of the interior-point method to enforce the irreversibility condition present in phase-field models for fracture. Consequently, the efficiency of the monolithic solver was only compared to the alternating algorithm when coupled with the interior-point method. The performances of the two solvers were concluded to be problem-dependent in this case. We also point out that few popular formulations, such as the history variable approach of Miehe and co-workers , or the hybrid model from , although computationally efficient, have the downfall of being variationally inconsistent. Gerasimov and De Lorenzis came to the conclusion that the approach based on the history variable proposed by Miehe leads to the solution of a problem that is not equivalent to the variational one . In this work, we formulate a fully monolithic and implicit solver relying on a modified Newton method with inertia correction of the Jacobian and a backtracking energy line-search, similar to . We then compare the performance of the alternating minimization algorithm, the quasi-monolithic scheme and the modified Newton solver on four benchmark tests selected from the phase-field literature. To eliminate the time step dependency of the quasi-monolithic scheme and facilitate its benchmarking, we couple the latter with an extrapolation correction loop controlled by a damage evolution . The benchmarking demonstrates that not only is the modified Newton method a robust monolithic solver for variational phase-field models, but it is also significantly more efficient than the quasi-monolithic and alternating minimization algorithms. The paper is organized as follows. In Section [2](#Background){reference-type="ref" reference="Background"}, we recall the formulation of variational phase-field models, describe the adopted treatment of the irreversibility constraint, and write the Euler-Lagrange equations of the minimization problem along with their discretization. We also present the alternating minimization algorithm and the quasi-monolithic scheme. In Section [3](#New solution algorithms){reference-type="ref" reference="New solution algorithms"}, we present a monolithic solver for variational phase-field models based on a modified Newton method and describe the quasi-monolithic scheme augmented with an extrapolation correction loop. Section [4](#Numerical results){reference-type="ref" reference="Numerical results"} compares the performance of the three studied algorithms on four different benchmark tests and discusses the results. Finally, our conclusions are presented in Section [5](#Conclusions){reference-type="ref" reference="Conclusions"}.
# Background {#Background}
## Variational modeling of brittle fracture
Here we define the phase-field model for brittle fracture, including the strain energy decomposition and treatment of the irreversibility condition adopted, on which is performed the benchmarking of the three solvers. Then, the Euler-Lagrange equations of the unconstrained variational problem and their spatial discretization are defined.
### Variational formulation
Let \(\Omega \subset {\rm I\!R}^\delta\) (\(\delta = 2,3\)) be the domain representing a crack-free brittle elastic body. The boundary \(\partial \Omega\) is subject to time-dependent Dirichlet and Neumann conditions, \(\Gamma_D \subset \partial \Omega\), \(\Gamma_N \subset \partial \Omega\), with \(\Gamma_D \cup \Gamma_N = \partial \Omega\) and \(\Gamma_D \cap \Gamma_N = \emptyset\). We also introduce an evolving crack surface \(\Gamma_c \subset \Omega\). We define the displacement function \(\boldsymbol{u}:\Omega \setminus \Gamma_c \subset {\rm I\!R}^\delta \rightarrow {\rm I\!R}^\delta\) and the linearized elastic strain \(\boldsymbol{e}(\boldsymbol{u})\coloneqq \frac{1}{2}(\nabla \boldsymbol{u}^T + \nabla \boldsymbol{u})\). We can therefore define the elastic potential energy \(\Psi(\boldsymbol{u}) = \frac{1}{2} \boldsymbol{\sigma}(\boldsymbol{u}):\boldsymbol{e}(\boldsymbol{u})\), where \(\boldsymbol{\sigma}(\boldsymbol{u})\) is the stress tensor. Finally, we denote the critical energy release rate of the brittle material with \(G_c\). Neglecting the external forces, Francfort and Marigo's of quasi-static brittle fracture leads to the following variational problem \[\boldsymbol{u},\Gamma_c = \textrm{arg}\,\textrm{min } \mathcal{E}(\boldsymbol{u},\Gamma_c ) \quad \textrm{s.t.} \quad \dot{\Gamma}_c \geq 0,\] with the energy functional \[\mathcal{E}(\boldsymbol{u}, \Gamma_c) = \int_{\Omega \setminus \Gamma_c} \Psi(\boldsymbol{u}) \, \text{d}\boldsymbol{x} + \int_{\Gamma_c} G_c \, \text{d}s.\] Phase-field damage models use the Ambrosio and Tortorelli elliptical regularization to overcome problems associated with the numerical treatment of the discontinuous crack \(\Gamma_c\), as proposed initially in . The regularization replaces the sharp crack by a diffuse damage band through the phase-field variable, denoted by \(d:\Omega \rightarrow [0,1]\), where \(d=0\) describes the undamaged material and \(d=1\) describes the fully damaged material. The displacement function becomes \(\boldsymbol{u}:\Omega \subset {\rm I\!R}^\delta \rightarrow {\rm I\!R}^\delta\). The different phase-field models for brittle fracture are characterized mainly by the type of crack function \(\alpha(d)\) and degradation function \(g(d)\) adopted. They can generally be expressed as \[\mathcal{E}_l(\boldsymbol{u},d) = \int_{\Omega} g(d)\Psi(\boldsymbol{u}) \, \text{d}\boldsymbol{x} + \frac{G_c}{c_\alpha}\int_{\Omega}\Big(\frac{\alpha(d)}{l}+l|\nabla d|^2 \Big)\text{d}\boldsymbol{x},\] with the normalization parameter \(c_\alpha \coloneqq 4\int_0^1 \sqrt{\alpha(z)} \, \text{d}z\). The regularization \(\alpha(d)\) and the characteristic length \(l\) define the shape and width of the smeared crack while the degradation function \(g(d)\) describes the dependency of the apparent stiffness on the damage field. In this paper, we focus on the so-called \(AT_1\) model with its linear crack function \(\alpha(d) = d\) and quadratic degradation function \(g(d)=(1-d)^2\) introduced by Pham et al. , yielding the functional \[\label{AT1} \mathcal{E}_l(\boldsymbol{u},d) = \int_{\Omega} (1-d)^2\Psi(\boldsymbol{u}) \, \text{d}\boldsymbol{x} + \frac{3G_c}{8}\int_{\Omega}\Big(\frac{d}{l}+l|\nabla d|^2 \Big)\text{d}\boldsymbol{x}.\] With the regularized energy functional relying on the phase-field variable, the variational problem now reads \[\label{VP} \boldsymbol{u},d = \textrm{arg}\,\textrm{min } \mathcal{E}_l(\boldsymbol{u},d ) \quad \textrm{s.t.} \quad 0 \leq d \leq 1 \quad \textrm{and} \quad \dot{d} \geq 0.\] As opposed to the widely used \(AT_2\) model, the \(AT_1\) model prevents damage initiation at low strain-energy density, therefore exhibiting an elastic phase. Furthermore, the linear crack function of the \(AT_1\) model leads to a finite localization band, implying the convergence to the Griffith energy without the necessity of the characteristic length \(l\) to approach zero . However, the constraint \(d \geq 0\) from the minimization problem [\[VP\]](#VP){reference-type="eqref" reference="VP"} is no longer naturally respected. A well-known issue with formulation [\[AT1\]](#AT1){reference-type="eqref" reference="AT1"} is its inability to distinguish between energy generated from tensile or compressive loading. Consequently, interpenetration of crack surfaces and non-physical cracks were observed early in the developments of the variational approach to fracture . Multiple methods using energy splits have been proposed to elude this difficulty , such as the volumetric-deviatoric split adopted in and the spectral split adopted in . In this work, we adopt the spectral decomposition of the strain tensor in its positive and negative components, as proposed by Miehe , with \[\boldsymbol{e}(\boldsymbol{u}) = \boldsymbol{e}^+(\boldsymbol{u}) + \boldsymbol{e}^-(\boldsymbol{u}),\] and \[\boldsymbol{e}^+(\boldsymbol{u}) = \boldsymbol{S}(\boldsymbol{u}) \boldsymbol{\Lambda}^+(\boldsymbol{u}) \boldsymbol{S}^T(\boldsymbol{u}),\] where \(\boldsymbol{\Lambda}^+\) is the diagonal matrix containing the positive eigenvalues \(\lambda_i^+\), and \(\boldsymbol{S}\) is the matrix containing the corresponding eigenvectors \(\boldsymbol{s}_i\) . Dropping the dependencies in \(\boldsymbol{u}\), the positive and negative stress of an isotropic material now read \[\boldsymbol{\sigma}^+ = \lambda \langle \text{tr}(\boldsymbol{e}) \rangle_{+} \boldsymbol{I} + 2\mu \boldsymbol{e}^+,\] \[\boldsymbol{\sigma}^-= \lambda \langle \text{tr}(\boldsymbol{e}) \rangle_{-} \boldsymbol{I} + 2\mu \boldsymbol{e}^-,\] where \(\boldsymbol{I}\) is the second order identity tensor and (\(\lambda\), \(\mu\)) are the Lamé coefficients. \(\langle z \rangle_{\pm}\) designates the positive or negative ramp function of \(z\), where \(\langle z \rangle_{+} = \text{max}(0,z)\) and \(\langle z \rangle_{-} = \text{min}(0,z)\). The energy functional using the energy split becomes \[\begin{split} \label{AT1+split} \mathcal{E}_l(\boldsymbol{u},d) = & \int_{\Omega} \frac{1}{2}\Big((1-d)^2\boldsymbol{\sigma}^+(\boldsymbol{u}) + \boldsymbol{\sigma}^-(\boldsymbol{u})\Big):\boldsymbol{e}(\boldsymbol{u}) \, \text{d}\boldsymbol{x} \\ & + \frac{3G_c}{8}\int_{\Omega}\Big(\frac{d}{l}+l|\nabla d|^2 \Big) \text{d}\boldsymbol{x}. \end{split}\]
### Irreversibility constraint {#irreversibility}
Formulation [\[AT1+split\]](#AT1+split){reference-type="eqref" reference="AT1+split"} is set in a quasi-static regime, with dependency on time appearing only through the boundary conditions. However, the irreversibility condition found in [\[VP\]](#VP){reference-type="eqref" reference="VP"} implies an history dependency in the minimization problem. Introducing a discretization of the loading process through the pseudo-time steps \(\Delta t^n = t^n-t^{n-1}\), and knowing the solution from the previous step, we look for a solution to the minimization problem at every time increment. Using the time discretization, \(\dot{d}\geq 0\) can be approximated through a backward differentiation with \[d^n \geq d^{n-1}.\] To avoid a cumbersome notation, we denote the solution of the actual increment \((\boldsymbol{u}^n,d^n)\) with \((\boldsymbol{u},d)\). The variational problem [\[VP\]](#VP){reference-type="eqref" reference="VP"} now contains 3 constraints, namely \(d \geq 0\), \(d \leq 1\) and \(d \geq d^{n-1}\). With the degradation function chosen as quadratic, \(d \leq 1\) is naturally respected . However, \(d \geq 0\) and \(d \geq d^{n-1}\) need to be imposed. Multiple options for enforcing the \(d \geq d^{n-1}\) condition are available in the literature, such as , the penalty method , , the augmented Lagrangian method , or the interior-point method . Here we adopt the simple quadratic penalty, as proposed by \[\label{penalty} P = \frac{\gamma}{2} \int_{\Omega} \langle d-d^{n-1} \rangle_{-}^{2}\: \text{d}\boldsymbol{x},\] where \(\gamma\) is a penalty parameter and \(d^{n-1}\) is the solution of the phase-field from the previous time increment. By taking \(d^0 \geq 0\), the constraint \(d \geq 0\) is simultaneously enforced. Adding the penalty term [\[penalty\]](#penalty){reference-type="eqref" reference="penalty"} to the energy functional [\[AT1+split\]](#AT1+split){reference-type="eqref" reference="AT1+split"} yields an unconstrained minimization problem which reads \[\label{VPu} \boldsymbol{u},d = \textrm{arg}\,\textrm{min } \mathcal{E}_l(\boldsymbol{u},d)\] with the energy functional \[\begin{split} \label{AT1+split+P} \mathcal{E}_l(\boldsymbol{u},d) = & \int_{\Omega} \frac{1}{2}\Big((1-d)^2\boldsymbol{\sigma}^+(\boldsymbol{u}) + \boldsymbol{\sigma}^-(\boldsymbol{u})\Big):\boldsymbol{e}(\boldsymbol{u}) \, \text{d}\boldsymbol{x} \\ & + \frac{3G_c}{8}\int_{\Omega}\Big(\frac{d}{l}+l|\nabla d|^2 \Big) \text{d}\boldsymbol{x} + \frac{\gamma}{2} \int_{\Omega} \langle d-d^{n-1} \rangle_{-}^{2}\: \text{d}\boldsymbol{x}. \end{split}\] This new problem is an approximation of [\[VP\]](#VP){reference-type="eqref" reference="VP"}, where the minimizers of [\[AT1+split+P\]](#AT1+split+P){reference-type="eqref" reference="AT1+split+P"} converge to the minimizers of [\[AT1+split\]](#AT1+split){reference-type="eqref" reference="AT1+split"} with \(\gamma \rightarrow \infty\). Instead of increasing \(\gamma\) until the Karush-Kuhn-Tucker (KKT) conditions are met, as one would typically do with a penalty method, we use the fixed optimal penalty parameter presented in \[\gamma = \frac{G_c}{l}\frac{27}{64 \mathtt{TOL^2_{Ir}}},\] where \(\mathtt{TOL_{Ir}}\) is a chosen irreversibility tolerance. . Derived from the notion of optimal phase-field profile, this constant penalty coefficient was shown to lead to a good approximation of the original constrained problem .
### Euler-Lagrange equations and finite element discretization {#section:ELd}
We define the function spaces and \(\mathcal{W} \coloneqq \hll{\{ w \in H^1(\Omega) \}}\), with \(H^1(\Omega)\) the Sobolev space. Applying directional derivatives towards \((\boldsymbol{u},d)\) to [\[AT1+split+P\]](#AT1+split+P){reference-type="eqref" reference="AT1+split+P"}, the Euler-Lagrange equations of problem [\[VPu\]](#VPu){reference-type="eqref" reference="VPu"} can be written in their weak form as :
The equation system [\[EL\]](#EL){reference-type="eqref" reference="EL"} is usually discretized using a Galerkin finite element method. To this end, assuming a 2D polygonal domain, we partition \(\Omega\) in a non-degenerate family of quadrilateral elements \(\mathcal{Q}^h\), where \(h\) is the average element edge size. Using the notation from , we introduce the linear finite element space \[\mathcal{V}_h \coloneqq \{ v \in H^1(\Omega): \text{ \(v|_Q\) is linear \(\forall Q \in \mathcal{Q}^h\)}\}.\] The nodal basis functions of the vector space \(\mathcal{V}_h\) are noted \[\{ \varphi_i: 1 \leq i \leq m \},\] where \(m\) is the dimension of the finite element space. To approximate the vector field \(\boldsymbol{u}\), and borrowing the deal.II convention , we define the \(2m\) vector shape functions \[\begin{aligned}
\boldsymbol{\phi}_0(\boldsymbol{x}) & = \begin{pmatrix} \varphi_0(\boldsymbol{x}) \\ 0 \end{pmatrix}, \\ \boldsymbol{\phi}_1(\boldsymbol{x}) & = \begin{pmatrix} 0 \\ \varphi_0(\boldsymbol{x}) \end{pmatrix}, \\ \boldsymbol{\phi}_2(\boldsymbol{x}) & = \begin{pmatrix} \varphi_1(\boldsymbol{x}) \\ 0 \end{pmatrix}\hll{,} \ldots \\ \boldsymbol{\phi}_{2m}(\boldsymbol{x}) & = \begin{pmatrix} 0 \\ \varphi_m(\boldsymbol{x}) \end{pmatrix}. \end{aligned}\] Collecting the displacement and damage nodal coefficients in respectively \(\boldsymbol{U} \in \mathbb{R}^{2m}\) and \(\boldsymbol{D} \in \mathbb{R}^m\), the functions \((\boldsymbol{u},d)\) are now approximated with \[\boldsymbol{u}_h(\boldsymbol{x}) = \sum_{j = 1}^{2m}\boldsymbol{\phi}_j(\boldsymbol{x})U_j, \quad d_h(\boldsymbol{x}) = \sum_{j = 1}^{m}\varphi_j(\boldsymbol{x}) D_j.\] Using the finite element approximations \(\boldsymbol{u}_h \in \mathcal{V}_h \times \mathcal{V}_h\) and \(d_h \in \mathcal{V}_h\), the discretized Euler-Lagrange equations can be written as the following residuals
For all Newton-based solvers, the assembly of the Jacobian, or at least part of it, is necessary. The Jacobian associated to [\[ELd\]](#ELd){reference-type="eqref" reference="ELd"} can be written in its block structure as
with the general structure \[\boldsymbol{J} = \begin{pmatrix} \boldsymbol{J}^{uu} & \boldsymbol{J}^{ud} \\ \boldsymbol{J}^{du} & \boldsymbol{J}^{dd} \end{pmatrix}.\] By \(\boldsymbol{\sigma}^+(\boldsymbol{\phi}_j)\), we imply the directional derivative of \(\boldsymbol{\sigma}^+(\boldsymbol{u}_h)\) towards \(U_j\) as thoroughly detailed by . Stemming from the derivative of the ramp function \(\langle z \rangle_{-}\), \(H^-(z)\) is the Heaviside function worth 1 if \(z \leq 0\) and worth 0 otherwise. Finally, we denote the discretized Dirichlet conditions \(\boldsymbol{\bar{U}}\).
## Solution algorithms {#Solution algorithms}
We present two algorithms commonly used to obtain a solution in phase-field fracture problems: the alternating minimization algorithm and the quasi-monolithic scheme. With these two methods, the Euler-Lagrange equations [\[EL\]](#EL){reference-type="eqref" reference="EL"} are treated as a set of non-linear equations and only a stationary point of [\[AT1+split+P\]](#AT1+split+P){reference-type="eqref" reference="AT1+split+P"} is obtained. The nature of the stationary point as either a local minimum, saddle point, or local maximum is unknown.
### Alternating minimization algorithm
As observed in , the functional [\[AT1\]](#AT1){reference-type="eqref" reference="AT1"} becomes convex in one field when the other is fixed. A stationary point can therefore be easily obtained with Newton's method if we solve towards one field at a time. Based on this fact, the alternating minimization algorithm, also often referred to as the staggered solver, consists in splitting the problem in a kinematic and a damage sub-problem and, as its name implies, alternatively solving them until convergence in both \(\boldsymbol{u}\) and \(d\). Since the solution is independent from the previous time steps, except of course for the penalty term, the alternating minimization algorithm is an implicit solver. Applied to the energy functional [\[AT1+split+P\]](#AT1+split+P){reference-type="eqref" reference="AT1+split+P"}, the staggered solver sequentially solves [\[ELd1\]](#ELd1){reference-type="eqref" reference="ELd1"} and [\[ELd2\]](#ELd2){reference-type="eqref" reference="ELd2"} using their respective Jacobian \(J^{uu}\) [\[J1\]](#J1){reference-type="eqref" reference="J1"} and \(J^{dd}\) [\[J4\]](#J4){reference-type="eqref" reference="J4"} through Newton's method. In this work, we adopt the alternating minimization algorithm as presented in , and described in Algorithm [\[Stagg\]](#Stagg){reference-type="ref" reference="Stagg"}.
## Single-edge notched plate under shear loading {#SENP-shear}
The notched square plate under shear loading (SENP-shear) is another widely adopted test in the phase-field community since it allows the verification of the energy splitting method . From the latter numerical experiments found in the literature, it appears that only reported results for the \(AT_1\) model. We used the same parameters to allow a comparison with the results from , with \(E = 210\) GPa, \(\nu = 0.3\), \(G_c = 2.7\) N/mm and \(l = 0.01\) mm. The geometry, the boundary conditions and the mesh are presented in Figure [\[fig:Shear\]](#fig:Shear){reference-type="ref" reference="fig:Shear"}. A total horizontal displacement of 0.015 mm is applied in 50 equal increments. The mesh is made of 23341 nodes assembled in 23205 elements. In this case, a smaller tolerance of \(\mathtt{TOL}_{QM} = 5\times 10^{-3}\) was required because a delay of the initiation and an oscillation of the reaction force were observed for a greater tolerance (results not shown). Figure [\[fig:CrackShear\]](#fig:CrackShear){reference-type="ref" reference="fig:CrackShear"} shows the crack paths obtained for the SENP-shear test. As expected and consistently with the results found in the literature (see for example ), the cracks initiate from the pre-crack and converge in a curved trajectory towards the lower right corner. The cracks yielded by the three solvers are virtually identical. Figure [\[fig:ShearFD\]](#fig:ShearFD){reference-type="ref" reference="fig:ShearFD"} presents the load-displacement curves obtained with the three solvers for the SENP-shear test. The responses of the three solvers coincide and match with that of . The difference in peak force observed between the results of Figure [\[fig:ShearFD\]](#fig:ShearFD){reference-type="ref" reference="fig:ShearFD"} and the results of could be explained by their use of \(P_1\) triangular elements and a mesh size of \(h = \frac{l}{4}\). The cumulative number of iterations required by the three solvers is presented in Figure [\[fig:ShearIT\]](#fig:ShearIT){reference-type="ref" reference="fig:ShearIT"}. The efficiency of the modified Newton method is clearly illustrated, with a total number of iterations an order of magnitude lower than the quasi-monolithic and alternating schemes. We also note that the iteration totals required by the alternating and modified Newton methods are of the same order of magnitude than that of . Table [2](#table:Shear){reference-type="ref" reference="table:Shear"} summarizes the computational performances of the three methods for the SENP-shear test. As presented, the total number of iterations required by the modified Newton solver is respectively 15 and 7 times lower when compared to the staggered and the extrapolated scheme. On our workstation, the iteration reduction meant a decrease in computation time by factors of respectively 9.25 and 6.40. Finally, one can observe from Figure [\[fig:ShearIT\]](#fig:ShearIT){reference-type="ref" reference="fig:ShearIT"} that the quasi-monolithic scheme is as efficient as the monolithic modified Newton method prior to the damage propagation. However, the efficiency significantly decreases once the crack starts propagating.
-2cm
## Three-point bending test
Let us now consider the classic three-point bending of a notched beam as presented in , . The geometry, the loading and the discretization used are depicted in Figure [\[fig:3Point\]](#fig:3Point){reference-type="ref" reference="fig:3Point"}. We adopt the same material properties as in , with \(E = 20.8\) GPa, \(\nu = 0.3\) and \(G_c = 0.54\) N/mm. It should be noted, however, that the simulations performed in were made using the \(AT_2\) model. In order to have equivalent mechanical properties for the \(AT_1\) model, the characteristic length is converted using the solution for a homogeneous bar under uniaxial tension . The conversion yields a characteristic length of \(l = 0.101\) mm. A total displacement of 0.1 mm discretized in 50 equal increments is applied as illustrated in Figure [\[fig:3Point\]](#fig:3Point){reference-type="ref" reference="fig:3Point"}. 11181 nodes and 11031 elements are used. As in the SENP-tension test, a tolerance of \(\mathtt{TOL}_{QM} = 0.01\) was found to be sufficient for the quasi-monolithic scheme to converge to the implicit solution. The resulting crack patterns for the three-point bending test are shown in Figure [\[fig:Crack3Point\]](#fig:Crack3Point){reference-type="ref" reference="fig:Crack3Point"}. As expected for a mode-I fracture, the cracks for all three solvers propagate from the notch to the upper edge in a straight line. Once again, they are virtually identical with only a negligible offset for the solution obtained through the modified Newton solver. Figure [\[fig:3PointFD\]](#fig:3PointFD){reference-type="ref" reference="fig:3PointFD"} presents the load-deflection curves obtained for the three-point bending test with the three methods. The solutions are identical and in good agreement with that reported in . Figure [\[fig:3PointIT\]](#fig:3PointIT){reference-type="ref" reference="fig:3PointIT"} presents the cumulative iterations required by the three solvers to obtain a solution to the phase-field problem. As illustrated, the quasi-monolithic and the modified-Newton methods have a similar efficiency before the propagation of damage. However, once the crack starts propagating, the modified Newton algorithm clearly outperforms the alternating and quasi-monolithic schemes. Table [3](#table:3Point){reference-type="ref" reference="table:3Point"} shows the total number of iterations, the maximum number of iterations in an increment and the calculation time required by the three algorithms. The modified Newton solver requires a maximum of only 193 iterations, while the extrapolated and alternating schemes require 565 and 989 iterations, respectively. This is a reduction in the number of iterations by factors of \(\approx 3\) and \(\approx 5\), which translates to an acceleration of the computation by factors of 3.33 and 4.18, respectively.
-2cm
## L-shaped panel test
The concrete L-shaped panel with mixed mode crack propagation from is another popular experiment used to evaluate and validate phase-field models . We adopt the geometry, loading and material properties from for comparison purposes. The geometry, boundary conditions and mesh are presented in Figure [\[fig:LShaped\]](#fig:LShaped){reference-type="ref" reference="fig:LShaped"}. The material parameters are \(E = 21.85\) GPa, \(\nu = 0.18\), \(G_c = 0.095\) N/mm and \(l = 3.125\) mm. A total vertical displacement of 1.0 mm is applied as described in Figure [\[fig:LShaped\]](#fig:LShaped){reference-type="ref" reference="fig:LShaped"}. 50 equal increments are used to discretize the loading process. The finite element mesh is made of 48984 nodes and 48850 quadrilateral elements. A tolerance of \(\mathtt{TOL}_{QM} = 5\times 10^{-3}\) is used for the results presented for the quasi-monolithic scheme, but as will be shown in the following, no convergence to the implicit solution was obtained. The crack paths obtained with the three solvers for the L-shaped panel tests are shown in Figure [\[fig:CrackLShaped\]](#fig:CrackLShaped){reference-type="ref" reference="fig:CrackLShaped"}. The crack patterns resulting from the alternating minimization and the modified Newton solvers are identical and in agreement with results found in the literature, e.g. , where the crack first propagate with an angle and stabilizes in the horizontal direction. A small deviation is observed at the end of the crack path but the results still concur with experimentally observed fractures . Figure [\[fig:LShapedFD\]](#fig:LShapedFD){reference-type="ref" reference="fig:LShapedFD"} shows the load-displacement curves obtained for the L-shaped panel with the three algorithms. Again here, the alternating minimization and modified Newton solvers are in agreement. The response also concurs with the results obtained without the research for a global minimum in . The difference in peak load can be explained by the absence of a tension/compression split and the crack initiation performed in . The crack and the load-displacement curve obtained with the quasi-monolithic scheme are in agreement with neither the numerical results obtained from the two other methods nor the experimental results reported in . Furthermore, the obtained crack path is physically questionable since the initiation angle does not agree with the maximum hoop stress criterion, stating that the crack should propagate in the direction maximizing the opening stress . This error can be attributed to the extrapolation correction loop, which in the present case appears to have significantly influenced the solution. However, either no convergence to the implicit solution or no convergence to a stationary point was obtained with different tolerances \(\mathtt{TOL}_{QM}\) or different uniform time refinements. Figure [\[fig:LShapedIT\]](#fig:LShapedIT){reference-type="ref" reference="fig:LShapedIT"} presents the performance of the three algorithms for the L-shaped panel test, although the results for the extrapolated scheme should be interpreted carefully since the method did not converge to the same solution. As can be seen, the final total iterations required by the modified Newton solver is lower than the alternating minimization algorithm by more than 10 times. Table [4](#table:LShaped){reference-type="ref" reference="table:LShaped"} summarizes the performances of the three methods for the L-shaped panel test. As noted in Table [4](#table:LShaped){reference-type="ref" reference="table:LShaped"}, a staggering number of iterations was required by the alternating minimization algorithm to solve increment 14 (\(\bar{u}_y = 0.28\) mm). In comparison, the modified Newton solver required 26 times less iterations for the same increment.
-2cm
## Discussion of the results
We recall that the alternating minimization algorithm, the quasi-monolithic scheme and the modified Newton method all used the same numerical implementation and treatment of the irreversibility condition. The difference remained in the algorithm used to obtain a stationary point or local minimum. The benchmarking on the four tests, selected from the phase-field literature, showed that both the extrapolated scheme and the modified Newton method offered an increase in computational efficiency, when compared to the staggered algorithm. However, once damage started propagating, the semi-implicit extrapolated scheme required additional correction to converge to the implicit solution. Consequently, the modified Newton method was found to be, by far, the most efficient solver. For example, on the SENP-shear test, the modified Newton solver required up to 11 and 3 times less iterations by increment than the alternating minimization and quasi-monolithic algorithms, respectively. For the L-shaped panel test, the increase in efficiency obtained with the modified Newton solver implied a decrease in computation time by a factor of 12 over the staggered solver. The alternating minimization algorithm remains a powerful and versatile method since it can be applied to non-variational models and implemented using a variational inequality solver, like reported in . The performance of the extrapolated scheme can probably be enhanced with locally refined time steps and a well-tuned extrapolation. It can also be applied to non-variational formulation. , convergence problems were encountered with the quasi-monolithic scheme, as noted for the L-shaped panel test. It is possible that the preconditioned GMRES solver reported in would improve the method. The numerical parameters appearing in the inertia correction algorithm of the modified Newton method could also be tuned. Finally, as already mentioned, convergence issues were sporadically observed with the quasi-monolithic scheme and a surprisingly high number of iterations were necessary for the three solvers on the L-shaped panel test. The latter issues are symptomatic of ill-conditioned non-linear systems and could be attributed to the quadratic penalty term used to enforce the irreversibility condition, even if the optimal penalty parameter from was used. As presented in Section [2.1.2](#irreversibility){reference-type="ref" reference="irreversibility"}, other methods of enforcing the irreversibility condition are available in the literature. Now that an implicit monolithic solver has been identified, the benchmarking of the schemes used to enforce the \(\dot{d} \geq 0\) constraint could allow the identification of a method reducing the issues associated to bad conditioning and minimizing the computational cost associated to the solution of phase-field models.
# Conclusion {#Conclusions}
A robust, efficient and straightforward solver is still lacking the variational phase-field fracture literature. In an attempt to fill this gap, we proposed a fully monolithic solver based on a modified Newton method with inertia correction and an energy line-search. Additionally, we augmented the quasi-monolithic algorithm with an extrapolation correction loop controlled by a damage-based criterion. Although not perfect, the modified quasi-monolithic scheme allowed us to compare its performance with the alternating algorithm and the modified Newton solver. Through the four test cases presented involving different modes of propagation and initiation, we showed the modified Newton solver to be robust and more efficient than the alternating minimization algorithm and the quasi-monolithic scheme. More precisely, the proposed monolithic solver was shown to be up to \(\approx 12\) and \(\approx 6\) times faster than the staggered and extrapolated schemes. Furthermore, the performance of the monolithic method did not appear to be problem-dependent, requiring no parameter adjustment, and showing consistent efficiency gains across the test cases. Additional work can still be done to enhance the performance of phase-field solvers. A study concerning the identification of numerical parameters maximizing the efficiency of the modified Newton solver could be performed. As suggested in the discussion, the benchmarking of the methods used to enforce the irreversibility condition could also help further reduce the computational cost of the phase-field method. However, the combination of the modified Newton solver with an adaptive mesh technique could already facilitate the application of phase-field models to more complex structure and with sophisticated constitutive laws. | {'timestamp': '2021-09-15T02:30:28', 'yymm': '2109', 'arxiv_id': '2109.05373', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05373'} |
null | null |
null | null |
# Introduction {#sec:intro}
A common theme in statistical inference and machine learning is to select a few relevant features or variables (among many) that are associated with the responses of interest. Controlling the expected proportion of falsely selected features (or the FDR ) for the variable selection problem has attracted much attention since the development of the *knockoff filter* procedure due to its flexibility (in constructing test statistics) as well as good finite-sample performances . In this procedure, a test statistic is computed for every feature of the model and the variables are selected by (data-dependent) thresholding the statistics according to the target FDR. The knockoff filter has been generalized to various settings such as the high-dimensional case (\(p>n\)) , beyond linear models (the model-X framework) , and deep learning-based procedure [@romano2019deep], with applications in a variety of domains [@lu2018deeppink; @fan2019ipad]. In this work, we aim to develop \((\epsilon,\delta)\)-differentially-private statistics that ensure (finite-sample) FDR control while being empirically powerful. Differential privacy is currently the standard framework concerning the privacy of individuals. The differential privacy of other variable selection methods like the Lasso and BH procedure are investigated in and, respectively. However, the privacy of the knockoff procedure hasn't been discussed in the literature yet. Related works are where they consider the differential privacy of the sample covariance and the least-squares estimator. Although our approach is similar, the work is different in the sense that we take care of the FDR control and the knockoff design which is a function of the data. We adopt the sensitivity analysis and output perturbation method and randomize the statistics by adding noise according to Gaussian and Laplace mechanism.
# Preliminaries {#sec:prob}
## Fixed-X Knockoff Filter
Let \({\bf y}\in \mathbb{R}^n\) denote a vector of responses, and \({\bf X}\in\mathbb{R}^{n\times p}\) a data matrix consisting of \(p\) (explanatory) variables and \(n\) samples. A canonical linear regression model is given by \[\begin{aligned}
{\bf y} = {\bf X} \boldsymbol \beta + {\bf w},\label{linear_model} \end{aligned}\] where \(\boldsymbol \beta = [\beta_1...,\beta_p]^\top\in \mathbb{R}^p\) is the unknown regression vector and the model assumptions are as follows.
We say the \(i\)-th variable is a null variable if \(\beta_i = 0\). The goal is to select the non-null variables (i.e. \(\beta_i \neq 0\)) without selecting too many nulls. For the linear model [\[linear_model\]](#linear_model){reference-type="eqref" reference="linear_model"}, let \(\hat{S}\subseteq\{1,2,\hdots,p\}\) denote the selected variables by some procedure. Then the FDR is defined as follows \[\text{FDR}=\mathbb{E}\Bigg[\frac{\#\{j:\beta_j=0\ \text{and}\ j\in\hat{S}\}}{\#\{j: j\in\hat{S}\}\vee 1}\Bigg],\] where the expectation is taken over the randomness of the error terms, \({\bf w}\). A selection rule controls the FDR at level \(q\) if its corresponding FDR is guaranteed to be at most \(q\). In the knockoff procedure, the FDR control is achieved by computing a vector of statistics \({\bf W}=(W_1, W_2,\hdots,W_p)^\top\) with components corresponding to each variable and selecting \(\{j:W_j\geq T(q,\Omega)\}\), where the threshold \(T\geq 0\) depends on the target FDR \(q\) and the set of computed statistics, i.e., \(\Omega=\{W_1,\hdots,W_p\}\). The main idea in computing the statistics is to use a fake copy of the design that preserve the correlation structure of \({\bf X}\) but destroys the relationship between \({\bf X}\) and \({\bf y}\). For example, if we shuffle (or permute) the samples in \({\bf X}\), we preserve the geometry of \({\bf X}\) but the relation with \({\bf y}\) would be diminished. The knockoff filter is based on a more elegant construction of such a fake design, namely, the knockoff design.
### Knockoff Design
Knockoff copies are constructed using only the original features and the subtle geometrical relation between the data matrix \({\bf X}\) and its knockoff \(\mathbf{\Tilde{X}}\in\mathbb{R}^{n\times p}\) makes them an appropriate tool for FDR control. Let \(\boldsymbol{\Sigma}={\bf X}^\top{\bf X}\) denote the Gram matrix of the original features. The relationship between \({\bf X}\) and \(\mathbf{\Tilde{X}}\) can be summarized as below \[\begin{aligned}
&{\bf G} := [ {\bf X}\ \mathbf{\Tilde{X}}]^\top[ {\bf X}\ \mathbf{\Tilde{X}}]\nonumber\\ &\:\quad=\begin{bmatrix} {\boldsymbol{\Sigma}\qquad\qquad\quad \boldsymbol{\Sigma}-\text{diag}\{{\bf s}\}}\\{\boldsymbol{\Sigma}-\text{diag}\{{\bf s}\}\qquad\qquad\quad \boldsymbol{\Sigma}}\end{bmatrix},\label{eq:data_gram} \end{aligned}\] where \({\bf s}\in\mathbb{R}^p_+\) is chosen in a way that \(\bf G\) is positive semidefinite. If \(n\geq 2p\), \(\mathbf{\Tilde{X}}\) can be constructed as follows \[\mathbf{\Tilde{X}}={\bf X}\,({\bf I}_p-\boldsymbol{\Sigma}^{-1} \text{diag}\{{\bf s}\})+\mathbf{\tilde{U}}{\bf C},\label{KOconst}\] where \(\bf C\) is obtained by Cholesky decomposition of the Schur complement of \({\bf G}\) and \(\mathbf{\tilde{U}}^{n\times p}\) is an orthonormal matrix that satisfies \(\mathbf{\tilde{U}}^\top{\bf X}=0\). The existence of \(\mathbf{\tilde{U}}\) is guaranteed by \(n\geq 2p\). By looking at [\[eq:data_gram\]](#eq:data_gram){reference-type="eqref" reference="eq:data_gram"}, one can observe that the knockoff features matrix \(\tilde{{\bf X}}\) preserves the correlation structure of \({\bf X}\) and the vector \({\bf s}\) gives us a degree of freedom to make knockoff features different than the original ones, e.g., if \({\bf s}=0\) then \(\mathbf{\Tilde{X}}={\bf X}\). For maximum detection power, the knockoff and original features should be as orthogonal to each other as possible. The restriction on this matter is imposed by \({\bf G}\succeq 0\) which is the necessary and sufficient condition for \(\mathbf{\Tilde{X}}\) to exist.
### Statistics
Using the knockoff features, we are now able to construct a vector of statistics \({\bf W}\) allowing FDR control. For instance, the Lasso penalized least-squares estimate can be used to construct the statistics. \[\boldsymbol{\hat{\beta}}^*:=\underset{{\bf b}\,\in\mathbb{R}^{2p}}{\arg\min}\Big\{\,\Big\|{\bf y}-[{\bf X}\ \mathbf{\Tilde{X}}]{\bf b}\Big\|_2^2\,+\lambda \|{\bf b}\|_1 \Big\},\label{eq:LS_model}\] where \(\lambda \geq 0\). Using this estimate, one way to define the statistics \({\bf W}\) is as follows \[W_i^{\textsc{LCD}} = |\hat{\beta}^*_i|-|\hat{\beta}^*_{i+p}|\ ,\qquad 1\leq i\leq p\.\label{stat:LCD}\] We can also define the statistics differently, \[W_i = \text{sgn}(|\hat{\beta}^*_i|-|\hat{\beta}^*_{i+p}|)\,\max(|\hat{\beta}^*_i|,|\hat{\beta}^*_{i+p}|)\.\label{lcsm}\] The key feature of these statistics that make them suitable for the purpose of FDR control is the *sign property* of the nulls, i.e., the signs of null statistics (signs of \(W=\{W_j: \beta_j=0\}\)) are independent of their magnitudes and have \(\mathrm{Bern}(1/2)\) distribution. We will look at this property in more details in the subsequent subsection. Unlike the nulls, non-null statistics are expected to get large positive values. Thus, a selection rule of the form \(\{j:W_j\geq T\}\) for some \(T>0\) seems reasonable. However, \(T\) should be characterized such that guarantees FDR control.
### IID Sign Property of Nulls
In general, satisfying the following two properties is sufficient to guarantee the sign property of the null statistics. Let \({\bf X}^{(i)}\) denote the \(i\)-th column of \({\bf X}\). ***Property I**: Swapping \({\bf X}^{(i)}\) and \(\mathbf{\Tilde{X}}^{(i)}\) for all \(i\in F\) and any \(F\subseteq \{1,2,\hdots,p\}\) has the effect of switching the signs of \(\{W_i:i\in F\}\). This property is called **antisymmetry** in.* In the least-squares problem [\[eq:LS_model\]](#eq:LS_model){reference-type="eqref" reference="eq:LS_model"}, it can be observed that swapping \({\bf X}^{(i)}\) and \(\mathbf{\Tilde{X}}^{(i)}\) for any \(i\in \{1,2,\hdots,p\}\) will swap \(\hat{\beta}^*_i\) and \(\hat{\beta}^*_{i+p}\) and this will result in switching the sign of \(W_i\) according to [\[stat:LCD\]](#stat:LCD){reference-type="eqref" reference="stat:LCD"} and [\[lcsm\]](#lcsm){reference-type="eqref" reference="lcsm"}. ***Property II**: Let \({\bf W}_{\text{swap}(F)}\) denote the vector of statistics obtained when \({\bf X}^{(i)}\) and \(\mathbf{\Tilde{X}}^{(i)}\) are swapped in the regression problem for all \(i\in F\) where \(F\subseteq \{j:\beta_j = 0\}\). Then \({\bf W}_{\text{swap}(F)}\) and \({\bf W}\) have the same distribution, i.e. \({\bf W}_{\text{swap}(F)} \overset{d}{=}\, {\bf W}\).* In the least-squares problem [\[eq:LS_model\]](#eq:LS_model){reference-type="eqref" reference="eq:LS_model"}, it can be shown that the estimated coefficients for null variables and their corresponding knockoff variable are exchangeable which immediately implies **Property II**. Using the sign property for the null statistics, it can be shown that the selection rule \(\hat{\mathcal{S}}(T)=\{j:W_j\geq T\}\) controls the FDR at level \(q\in[0,1]\), where \[T = \min\Big\{t \in \Psi:\frac{1+\#\{j: W_j \leq-t\}}{\#\{j: W_j\geq t\}\vee 1}\leq q\Big\},\label{threshold}\] if \(\Psi =\{|W_i|:i=1,2,\hdots p\}\setminus\{0\}\) is not empty and \(T=+\infty\) if \(\Psi = \emptyset\).
## Differential Privacy
Differential privacy provides a rigorous formulation of individual privacy that allows for the development of (randomizing) mechanisms that release some information about the data without revealing too much about any individuals. In other words, the goal is to have a guarantee that the effect of each observation in the output is negligible.
The algorithms we are going to propose are based on the *output perturbation* mechanisms. In general, the class of additive noise mechanisms perturb the output by adding random noise proportional to the (differential) sensitivity of the output, that is, \(\mathcal{M}({\bf A}) = f({\bf A}) + Z\), where \(f\) is some (vector-valued) mapping and \(Z\) denotes the noise. The classical Gaussian mechanism and Laplace mechanism adopt \(Z\sim \mathcal{N}(0,\kappa^2 I)\) and \(Z_i\overset{\iid}{\sim}\text{Lap}(\eta)\), where \(\kappa^2\) and \(\eta\) are determined by the \(\ell_2\) and \(\ell_1\)-sensitivity of \(f\), respectively.
# Main Results
In this section we provide two methods for computing differential private statistic under the assumptions 1,2, and the following assumption.
For the purpose of variable selection with knockoffs, it is natural to normalize \({\bf X}\) by columns. So in both methods, the normalization is considered as a part of the procedure. Also, we let \(\text{diag}\{{\bf s}\} = \lambda_{\text{min}}(\boldsymbol \Sigma')\,{\bf I}_p\) in creating knockoffs [\[KOconst\]](#KOconst){reference-type="eqref" reference="KOconst"}, where \(\boldsymbol \Sigma' = {\bf X}'^\top{\bf X}'\) and \({\bf X}'\) denotes the design matrix \({\bf X}\) normalized by the \(\ell_2\)-norm of its columns.
## Method I
In this method we consider statistics that depend on the data through the following form, \[\widehat{{\bf W}}_1 = {\bf f}([{\bf X}'\ \widetilde{{\bf X}'}]^\top[{\bf X}'\ \widetilde{{\bf X}'}]+E,[{\bf X}'\ \widetilde{{\bf X}'}]^\top{\bf y}+{\bf e})\ ,\label{mthd1}\] where \(E\) and \({\bf e}\) denote the additive noise terms ensuring the differential privacy. Define \(C_{\text{min}}=\underset{1\leq i \leq p}{\text{min}}\big\|{\bf X}^{(i)}\big\|_2\) and \(\eta^2= \frac{B^2}{C_{\text{min}}^2-B^2}\), where \({\bf X}^{(i)}\) denotes the \(i\)-th column of \({\bf X}\). In the first argument of [\[mthd1\]](#mthd1){reference-type="eqref" reference="mthd1"}, \(E\in\mathbb{R}^{2p\times2p}\) is a random term of the following structure, \[E = \boldsymbol \theta_1 \begin{bmatrix}{{\bf O}_p\qquad{\bf I}_p}\\{{\bf I}_p\qquad {\bf O}_p}\end{bmatrix}+\begin{bmatrix}{1\quad 1}\\{1\quad 1} \end{bmatrix}\otimes\boldsymbol \theta_2 \nonumber\,\] where \(\boldsymbol \theta_1\sim\text{Lap}\big(\eta^2(1+ \lambda_{\text{min}}(\boldsymbol \Sigma'))/\epsilon_1\big)\) and \(\boldsymbol \theta_2\in\mathbb{R}^{p\times p}\) is a symmetric matrix with zero diagonal entries such that the elements of the upper triangle are drawn i.i.d. from \(\mathcal{N}(0,\kappa_1^2)\), where \(\kappa_1^2 \geq \frac{2\ln{(1.25/\delta)}}{\epsilon_2^2}\eta^4(\sqrt{2}+\|\boldsymbol \Sigma'\|_F)^2\), and each lower triangle entry is copied from its upper triangle counterpart. Regarding the second argument of [\[mthd1\]](#mthd1){reference-type="eqref" reference="mthd1"}, \({\bf e}\) is a vector with \(\mathcal{N}(0,\kappa_2^2)\) components with \(\kappa_2^2>\frac{2\ln{(1.25/\delta_1)}}{\epsilon^2}\Delta_2\big([{\bf X}'\ \widetilde{{\bf X}'}]^\top{\bf y}\big)^2\) computed according to the following \(\ell_2\)-sensitivity, \[\begin{aligned}
\Delta_2\big([{\bf X}'\ \widetilde{{\bf X}'}]^\top&{\bf y}\big) =\, \label{mthd1_Sen} \zeta^{\frac{1}{2}}\Big(2\sqrt{\gamma}+\eta\,\big(3+2\lambda_{\text{max}}(\boldsymbol \Sigma')\\+\lambda_{\text{min}}(\boldsymbol \Sigma')\big)^{\frac{1}{2}}&\Big)+\big\|\boldsymbol \beta\big\|\Bigg(\sqrt{2}\, \Big(\frac{\eta}{B}-\frac{1}{C_{\text{min}}}\Big)\big\|\boldsymbol \Sigma\big\|_F+2\,\eta B \nonumber\\ +\big(C_{\text{min}}-\frac{B}{\eta}\big)&\lambda_{\text{min}}(\boldsymbol \Sigma') +\eta^2\big(\lambda_{\text{min}}(\boldsymbol \Sigma')+1\big)\sqrt{C_{\text{min}}^2+B^2}\Bigg)\nonumber\, \end{aligned}\] where, \(\zeta = \frac{2p\,\sigma^2}{1-\sqrt{\frac{2}{p}\ln(2/\delta_2)}}\), \(\gamma=2\lambda_{\text{max}}(\boldsymbol \Sigma')-\lambda_{\text{min}}(\boldsymbol \Sigma')\), and \(\delta_2>2e^{-p/2}\).
For instance, we observe that if \({\bf f}({\bf A},{\bf b})\) denotes the vector of statistics computed using the estimate \({\bf A}^{-1}{\bf b}\), then \({\bf f}(\cdot,\cdot)\) satisfies the antisymmetry condition in Theorem [\[M1fdr\]](#M1fdr){reference-type="ref" reference="M1fdr"}, because we have \((P_F{\bf A} P_F)^{-1}P_F{\bf b}=P_F{\bf A}^{-1}{\bf b}\).
## Method II
In this method, we add noise directly to the least-squares estimate \(\boldsymbol{\hat{\beta}^{'}}\), i.e. [\[eq:LS_model\]](#eq:LS_model){reference-type="eqref" reference="eq:LS_model"} with normalized design and \(\lambda = 0\).
# Simulation Results {#sec:simulation}
In this section we present simulation results on synthetic data sets for both Method I and II. The sample size \(n\) is swept between \(10^3\) to \(10^{6.6}\). We have \(p = 50\) measurements per sample and the entries of the design matrix \({\bf X}\) are generated according to \(\mathcal{N}(0,{\bf I}_p)\). The responses \({\bf y}\) are generated according to the linear model [\[linear_model\]](#linear_model){reference-type="eqref" reference="linear_model"} with noise variance \(\sigma^2 = 1\) and \(k=15\) (non-zero) underlying coefficients of the same amplitude \(A=4.5\) and sign. We adopt the OLS estimator and the *Coefficient Signed Max* (CSM) statistic [\[lcsm\]](#lcsm){reference-type="eqref" reference="lcsm"}. The target FDR is \(q=0.2\) and the plots are based on averaging 250 trials. Each trial is \((0.2,\frac{2p}{n})\)-differentially private, meaning we let the overall \(\delta\) parameter drop as the sample size grows, so the privacy hold in a more strict sense for larger sample sizes[^1]. In Method I we set \(\epsilon = 0.1\), \(\epsilon_1=\epsilon_2=0.05\), and \(\delta=\delta_1=\delta_2\). In Method II we have \(\epsilon = 0.2\) and \(\delta_1=\delta_2\). The power is defined as follows, \[\text{Power}=\frac{1}{k}\,\mathbb{E}\big[\#\{j:\beta_j\neq 0\ \text{and}\ j\in\hat{S}\}\big]\.\]
[^1]: In practice it is common to take \(\delta\) to scale as \(\frac{1}{n}\). | {'timestamp': '2022-02-01T02:35:38', 'yymm': '2109', 'arxiv_id': '2109.05402', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05402'} |
# Introduction
Rule-based reasoning is an essential part of human intelligence prominently formalized in artificial intelligence research via logic programs with important applications to expert systems, database theory, and knowledge representation and reasoning (cf. ). Describing complex objects as the composition of elementary ones is a common strategy in computer science and science in general. has recently introduced the sequential composition of logic programs in the context of logic-based analogical reasoning and learning in logic programming. The sequential composition operation has been studied by in the propositional case and by in the non-monotonic case of answer set programming. Motivated by the aforementioned applications to analogical reasoning, in this paper we study the sequential decomposition of logic programs and show that it gives rise to a qualitative and algebraic notion of syntactic logic program similarity. More precisely, we say that a program \(P\) can be *one-step reduced* to a program \(R\)---in symbols, \(P\lesssim R\)---iff \(P=(Q\circ R)\circ S\) for some programs \(Q\) (prefix) and \(S\) (suffix). Now given a query \(q\) to \(P\), we can answer \(q\) by translating every SLD-resolution step of \(P\) via the prefix \(Q\) into an SLD-resolution step of \(R\) and translating the result back again with the suffix \(S\) (cf. ). We then say that \(P\) and \(R\) are *syntactically similar* iff \(P\) can be one-step reduced to \(R\) and vice versa, that is, iff \(P\lesssim R\) and \(R\lesssim P\). This definition has appealing mathematical properties. For example, we can show that the programs \(Plus\) for the addition of numerals and the program \(Append\) for list concatenation are syntactically similar according to our definition (cf. ). Interestingly, when translating queries to \(Append\) into queries of the seemingly simpler program \(Plus\), we obtain 'entangled' terms of the form \(s([\;\,])\) and \(s([b,c])\) which are neither numerals nor lists, and we believe that dual syntactic objects like these are essential for cross-domain analogical reasoning (see the discussion in ). Syntactic similarity is related to the number of bound variables per rule of a program and we deduce from this fact that, for example, the program \(Member\) for checking list membership, containing maximally two bound variables per rule, can be one-step reduced to the program \(Append\) containing maximally three bound variables per rule, but not vice versa: \(Member<Append\) (cf. Examples [\[exa:Member\]](#exa:Member){reference-type="ref" reference="exa:Member"} and [\[exa:Member2\]](#exa:Member2){reference-type="ref" reference="exa:Member2"}). This is interesting as the prefix \(Q\) witnessing \(Member=Q\circ Append\) resembles the mapping between \(Member\) and \(Append\) of introduced for analogical reasoning. In a broader sense, this paper is a further step towards an algebra of logic programs first envisioned by with applications to logic-based analogical reasoning and learning in logic programming.
# Logic Programs
We recall the syntax and semantics of logic programs by mainly following the lines of .
## Syntax {#sec:Syntax}
An (*unranked first-order*) *language* \(L\) is defined as usual from predicate, function, and constant symbols and variables. Terms and atoms over \(L\) are defined in the usual way. We denote the set of all ground \(L\)-atoms by \(HB_L\) or simply by \(HB\) called the *Herbrand base* over \(L\), and we denote the set of all \(L\)-atoms not containing function or constant symbols by \(XHB_L\). Substitutions and (most general) unifiers of terms and (sets of) atoms are defined as usual. We call any bijective substitution mapping variables to variables a *renaming*. Let \(L\) be a language. A (*Horn logic*) *program* over \(L\) is a set of *rules* of the form \[\begin{aligned}
\label{equ:r} A_0\leftarrow A_1,\ldots,A_k,\quad k\geq 0, \end{aligned}\] where \(A_0,\ldots,A_k\) are \(L\)-atoms. It will be convenient to define, for a rule \(r\) of the form , \(head(r):=\{A_0\}\) and \(body(r):=\{A_1,\ldots,A_k\}\), extended to programs by \(head(P):=\bigcup_{r\in P}head(r)\) and \(body(P):=\bigcup_{r\in P}body(r)\). In this case, the *size* of \(r\) is \(k\) denoted by \(sz(r)\). A *fact* is a rule with empty body and a *proper rule* is a rule which is not a fact. We denote the facts and proper rules in \(P\) by \(facts(P)\) and \(proper(P)\), respectively. A program \(P\) is *ground* if it contains no variables and we denote the grounding of \(P\) which contains all ground instances of the rules in \(P\) by \(gnd(P)\). We call a ground program *propositional* if it contains only propositional atoms with no arguments. The set of all *variants* of \(P\) is defined by \(variants(P):=\bigcup_{\theta\text{ renaming}}P[\theta]\). The variants of a program will be needed to avoid clashes of variables in the definition of composition. A variable is *bound* in a rule if it appears in the head and body. Define the *dual* of \(P\) by \[\begin{aligned}
P^d:=facts(P)\cup\{A\leftarrow head(r)\mid r\in proper(P): A\in body(r)\}. \end{aligned}\] Roughly, we obtain the dual of a theory by reversing all the arrows of its proper rules. We define the *left* and *right reduct* of \(P\), with respect to some interpretation \(I\), respectively by \[\begin{aligned}
^IP:=\{r\in P\mid I\models head(r)\} \quad\text{and}\quad P^I:=\{r\in P\mid I\models body(r)\}. \end{aligned}\]
## Semantics
An *interpretation* is any set of atoms from \(A\). We define the *entailment relation*, for every interpretation \(I\), inductively as follows: (i) for an atom \(a\), \(I\models a\) if \(a\in I\); (ii) for a set of atoms \(B\), \(I\models B\) if \(B\subseteq I\); (iii) for a rule \(r\) of the form , \(I\models r\) if \(I\models body(r)\) implies \(I\models head(r)\); and, finally, (iv) for a program \(P\), \(I\models P\) if \(I\models r\) holds for each rule \(r\in P\). In case \(I\models P\), we call \(I\) a *model* of \(P\). The set of all models of \(P\) has a least element with respect to set inclusion called the *least model* of \(P\). We say that \(P\) and \(R\) are *logically equivalent* if their least models coincide. Define the *van Emden-Kowalski operator* of \(P\), for every interpretation \(I\), by \(T_P(I):=\{head(r)\mid r\in P:I\models body(r)\}\). It is well-known that an interpretation \(I\) is a model of \(P\) iff \(I\) is a prefixed point of \(T_P\) and the least model of \(P\) coincides with the least fixed point of \(T_P\).
## SLD-Resolution
Logic programs compute via a restricted form of resolution, called *SLD-resolution*, as follows. For simplicity, we consider here only the ground case. Let \(q\) be a ground query \(\leftarrow A_1,\ldots,A_k\), \(k\geq 1\), and suppose that for some \(i\), \(1\leq i\leq k\) and \(m\geq 0\), \(r=A_i\leftarrow A'_1,\ldots,A'_m\) is a rule from \(gnd(P)\). Then \(q'\) given by \[\begin{aligned}
\leftarrow A_1,\ldots,A_{i-1},A'_1,\ldots,A'_m,A_{i+1},\ldots,A_k \end{aligned}\] is called a *resolvent* of \(q\) and \(r\), and \(A_i\) is called the *selected atom* of \(q\). By iterating this process we obtain a sequence of resolvents which is called an *SLD-derivation*. A derivation can be finite or infinite. If its last query is empty then we speak of an *SLD-refutation* of the original query \(q\). In this case we have derived an *SLD-proof* of \(A_1,\ldots,A_k\). A *failed* SLD-derivation is a finite derivation which is not a refutation. In case \(A\) is a ground atom with an SLD-proof from \(P\), we say that \(A\) is an *SLD-consequence* of \(P\) and write \(P\vdash A\). For a rule \(r\) of the form , we write \(P\vdash r\) in case \(P\vdash A_0\) whenever \(P\vdash A_i\) holds for every \(1\leq i\leq k\). We denote the *empty query* by \(\square\).
# Composition
In this section, we recall the sequential composition of logic programs as defined by and studied in the propositional case by .
The rule-like structure of logic programs induces naturally a compositional structure as follows.
Roughly, we obtain the composition of \(P\) and \(R\) by resolving all body atoms in \(P\) with the 'matching' rule heads of \(R\). This is illustrated in the next example, where we construct the even from the natural numbers via composition.
Notice that we can reformulate sequential composition as \[\begin{aligned}
\label{equ:bigcup} P\circ R=\bigcup_{r\in P}(\{r\}\circ R), \end{aligned}\] which directly implies right-distributivity of composition, that is, \[\begin{aligned}
\label{equ:(P_cup_Q)_circ_R} (P\cup Q)\circ R=(P\circ R)\cup (Q\circ R)\quad\text{holds for all propositional Horn theories }P,Q,R. \end{aligned}\] However, the following counter-example shows that left-distributivity fails in general: \[\begin{aligned}
\{a\leftarrow b,c\}\circ(\{b\}\cup\{c\})=\{a\} \quad\text{and}\quad (\{a\leftarrow b,c\}\circ\{b\})\cup(\{a\leftarrow b,c\}\circ\{c\})=\emptyset. \end{aligned}\] Define the *unit program* by the Krom program[^2] \[\begin{aligned}
1_L:=\{A\leftarrow A\mid A\in XHB_L\}. \end{aligned}\] In the sequel, we will omit the reference to \(L\). The space of all logic programs over some fixed language is closed under sequential composition with the neutral element given by the unit program and the empty program serves as a left zero, that is, we have \[\begin{aligned}
P\circ 1=1\circ P=1 \quad\text{and}\quad \emptyset\circ P=\emptyset. \end{aligned}\] We can simulate the van Emden-Kowalski operator on a syntactic level without any explicit reference to operators via sequential composition, that is, for any interpretation \(I\), we have \[\begin{aligned}
\label{equ:T_P} T_P(I)=gnd(P)\circ I. \end{aligned}\] As facts are preserved by composition and since we cannot add body atoms to facts via composition on the right, we have \[\begin{aligned}
\label{equ:IP=I} I\circ P=I. \end{aligned}\]
## Ground Programs
Ground programs are possibly infinite propositional programs where ground atoms can have a fixed inner structure, which means that we can import here the results of which do not depend on finiteness.
For ground programs, sequential composition simplifies to \[\begin{aligned}
P\circ R=\{head(r)\leftarrow body(S)\mid r\in P,S\subseteq_{sz(r)}R:head(S)=body(r)\}. \end{aligned}\] Our first observation is that we can compute the heads and bodies of a ground program \(P\) via \[\begin{aligned}
\label{equ:head(P)} head(P)=P\circ HB \quad\text{and}\quad body(P)=proper(P)^d\circ HB. \end{aligned}\] Moreover, we have \[\begin{aligned}
\label{equ:head(PR)} head(PR)\subseteq head(P) \quad\text{and}\quad body(PR)\subseteq body(R). \end{aligned}\] Given an interpretation \(I\), we define \[\begin{aligned}
I^\ominus:=1^{HB-I}\cup I \quad\text{and}\quad I^\oplus:=\{A\leftarrow(\{A\}\cup I)\mid A\in HB\}. \end{aligned}\] It is not difficult to show that \(PI^\ominus\) is the program \(P\) where all occurrences of the ground atoms in \(I\) are removed from the rule bodies in \(P\), that is, we have \[\begin{aligned}
PI^\ominus&=\{head(r)\leftarrow (body(r)-I)\mid r\in P\}. \end{aligned}\] Similarly, \(PI^\oplus\) is the program \(P\) with the ground atoms in \(I\) added to the rule bodies of all *proper* rules in \(P\), that is, we have \[\begin{aligned}
PI^\oplus&=facts(P)\cup\{head(r)\leftarrow (body(r)\cup I)\mid r\in proper(P)\}. \end{aligned}\] The left and right reducts can be represented via composition and the unit program by \[\begin{aligned}
\label{equ:^IP} ^IP=1^I\circ P \quad\text{and}\quad P^I=P\circ 1^I. \end{aligned}\] Interestingly enough, we can represent the grounding of a (non-ground) program \(P\) via composition with the grounding of the unit program by \[\begin{aligned}
\label{equ:gnd(P)} gnd(P)=(gnd(1)\circ P)\circ gnd(1). \end{aligned}\]
# Decomposition and Similarity
Let us start this section with a definition of a qualitative and algebraic notion of syntactic logic program similarity in terms of sequential decomposition.
The computational interpretation of \(P=(QR)S\) is that we can answer a query to \(P\) by translating it via \(Q\) into a query of \(R\), computing one SLD-resolution step with \(R\), and translating it back to \(P\) via \(S\) as is demonstrated in the next example.
# Properties of Similarity
Given ground atoms \(A_0,\ldots,A_k\), the identity \[\begin{aligned}
\{A_0\}=\{A_0\leftarrow A_1\}\{A_1\} \end{aligned}\] shows \[\begin{aligned}
\{A_0\}\lesssim \{A_0\leftarrow A_1\}. \end{aligned}\] Since we cannot add body atoms to facts via composition on the right, we have \[\begin{aligned}
\{A_0\leftarrow A_1\}\not\lesssim\{A_0\}. \end{aligned}\] Hence, \[\begin{aligned}
\{A_0\}<\{A_0\leftarrow A_1\}. \end{aligned}\] On the other hand, we have \[\begin{aligned}
\{A_0\leftarrow A_1\}&=\{A_0\leftarrow A_1,\ldots,A_k\}\{A_2,\ldots,A_k\}^\ominus,\\ \{A_0\leftarrow A_1,\ldots,A_k\}&=\{A_0\leftarrow A_1\}\{A_2,\ldots,A_k\}^\oplus, \end{aligned}\] which shows \[\begin{aligned}
\{A_0\leftarrow A_1\}\approx\{A_0\leftarrow A_1,\ldots,A_k\}. \end{aligned}\] In the non-ground case, we have, for example: \[\begin{aligned}
\{p\leftarrow p\}<\{p(x_1)\leftarrow p(x_1)\}<\{p(x_1,x_2)\leftarrow p(x_1,x_2)\}<\ldots. \end{aligned}\] This motivates the following definition.
The number of bound variables cannot increase via composition---for example: \[\begin{aligned}
(\{p(x,y)\leftarrow p(x)\}\circ\{p(x)\leftarrow p(x)\})\circ\{p(x)\leftarrow p(x,y)\}=\{p(x,y)\leftarrow p(x,z)\}. \end{aligned}\] This entails \[\begin{aligned}
width(PR)\leq width(P) \quad\text{and}\quad width(PR)\leq width(R). \end{aligned}\] Hence, we have \[\begin{aligned}
\label{equ:width} P\lesssim R \quad\Rightarrow\quad width(P)\leq width(R). \end{aligned}\]
The following propositions summarize some facts about syntactic similarity.
The following characterization of syntactic similarity follows immediately from .
The following simple example shows that syntactic similarity and logical equivalence are 'orthogonal' concepts.
# Conclusion {#sec:Conclusion}
This paper showed how a qualitative and algebraic notion of syntactic logic program similarity can be constructed from the sequential decomposition of programs. We derived some elementary properties of syntactic similarity and, more importantly, demonstrated how it can be used to answer queries across different domains by giving some illustrative examples. Interestingly, in , in the process of answering queries in the list domain by translating it to the seemingly simpler domain of numerals, we obtained 'entangled' terms of the form \(s([\;\,])\) and \(s([b,c])\) which are neither numerals nor lists. We believe that 'dual' syntactic objects of this form---which under the conventional doctrine of programming are not well-typed (see, for example, the remarks in )---are characteristic for 'creative' analogies across different domains and deserve special attention. More precisely, if we think of (logic) programs as tools for solving problems, then using those tools 'creatively' often requires, as in the example mentioned above, a tight coupling between seemingly incompatible objects. However, if those objects are statically typed by the programmer, then this might prevent the creation and exploitation of useful analogical 'bridges' for knowledge transfer.
[^1]: We write \(X\subseteq_k Y\) in case \(X\) is a subset of \(Y\) consisting of \(k\) elements.
[^2]: Recall from that \(XHB_L\) consists of all \(L\)-atoms not containing function or constant symbols. | {'timestamp': '2021-09-14T02:13:47', 'yymm': '2109', 'arxiv_id': '2109.05300', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05300'} |
# Introduction
Natural language processing (NLP) has long recognised morphosyntactic features as necessary for solving advanced natural language understanding (NLU) tasks. An enormous impact of contextual language models on presumably all NLP tasks has slightly weakened the importance of morphosyntactic analysis. As morphosytnactic features are encoded to some extent in contextual word embeddings, doubts arise as to whether explicit morphosyntactic knowledge is still needed. For example, have recently investigated an intermediate fine-tuning contextual language models on the dependency parsing task and suggested that this step does not significantly contribute to advance NLU models. Conversely, reveal the powerlessness of contextual language models in encoding linguistic phenomena like negation. This is in line with our intuition about representing negation in Polish sentences (see Figure [\[fig:trees\]](#fig:trees){reference-type="ref" reference="fig:trees"}). It does not seem trivial to differentiate between the contradicting meanings of these sentences using contextual language models, as the word context is similar. The morphosyntactic features, e.g. parts of speech PART vs. INTJ, and dependency labels *advmod:neg* vs. *discourse:intj*, could be beneficial in determining correct reading. In order to verify the influence of explicit morphosyntactic knowledge on NLU tasks, it is necessary to design a technique for injecting this knowledge into models or to build morphosyntax-aware representations. The first research direction was initiated by. Our objective is to provide a tool for predicting high-quality morphosyntactic features and exposing their embeddings. These vectors can be directly combined with contextual word embeddings to build morphosyntactically informed word representations. The emergence of publicly available NLP datasets, e.g. Universal Dependencies, stimulates the development of NLP systems. Some of them are optimised for efficiency, e.g. spaCy, and other for accuracy, e.g. UDPipe, the Stanford system, Stanza. In this paper, we introduce COMBO, an open-source fully neural NLP system which is optimised for both training efficiency and prediction quality. Due to its end-to-end architecture, which is an innovation within morphosyntactic analysers, COMBO is faster in training than the SOTA pipeline-based systems, e.g. Stanza. As a result of applying modern NLP solutions (e.g. contextualised word embeddings), it qualitatively outperforms other systems. COMBO analyses tokenised sentences and predicts morphosyntactic features of tokens (i.e. parts of speech, morphological features, and lemmata) and syntactic structures of sentences (i.e. dependency trees and enhanced dependency graphs). At the same time, its module, COMBO-vectoriser, extracts vector representations of the predicted features from hidden layers of individual predictors. COMBO user guide is in §[4](#sec:getting_started){reference-type="ref" reference="sec:getting_started"} and a live demo is available on the website <http://combo-demo.nlp.ipipan.waw.pl>.
# COMBO Architecture {#sec:architecture}
COMBO's architecture (see Figure [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"}) is based on the forerunner implemented in the Keras framework. Apart from a new implementation in the PyTorch library, the novelties are the BERT-based encoder, the EUD prediction module, and COMBO-vectoriser extracting embeddings of UPOS and DEPREL from the last hidden layers of COMBO's tagging and dependency parsing module, respectively. This section provides an overview of COMBO's modules. Implementation details are in Appendix [\[sec:implementationDetails\]](#sec:implementationDetails){reference-type="ref" reference="sec:implementationDetails"}.
#### Local Feature Extractors {#sec:localFeatureExtraction}
Local feature extractors (see Figure [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"}) encode categorical features (i.e. words, parts of speech, morphological features, lemmata) into vectors. The feature bundle is configurable and limited by the requirements set for COMBO. For instance, if we train only a dependency parser, the following features can be input to COMBO: internal character-based word embeddings ([char]{.smallcaps}), pre-trained word embeddings ([word]{.smallcaps}), and embeddings of lemmata ([lemma]{.smallcaps}), parts of speech ([upos]{.smallcaps}) and morphological features ([ufeats]{.smallcaps}). If we train a morphosyntactic analyser (i.e. tagger, lemmatiser and parser), internal word embeddings ([char]{.smallcaps}) and pre-trained word embeddings ([word]{.smallcaps}), if available, are input to COMBO. Words and lemmata are always encoded using character-based word embeddings ([char]{.smallcaps} and [lemma]{.smallcaps}) estimated during system training with a dilated convolutional neural network (CNN) encoder. Additionally, words can be represented using pre-trained word embeddings ([word]{.smallcaps}), e.g. fastText, or BERT. The use of pre-trained embeddings is an optional functionality of the system configuration. COMBO freezes pre-trained embeddings (i.e. no fine-tuning) and uses their transformations, i.e. embeddings are transformed by a single fully connected (FC) layer. Part-of-speech and morphological embeddings ([upos]{.smallcaps} and [ufeats]{.smallcaps}) are estimated during system training. Since more than one morphological feature can attribute a word, embeddings of all possible features are estimated and averaged to build a final morphological representation.
#### Global Feature Encoder {#sec:globalFeatureExtraction}
The encoder uses concatenations of local feature embeddings. A sequence of these vectors representing all the words in a sentence is processed by a bidirectional LSTM. The network learns the context of each word and encodes its global (contextualised) features (see Figure [\[fig:global\]](#fig:global){reference-type="ref" reference="fig:global"}). Global feature embeddings are input to the prediction modules.
#### Tagging Module {#sec:tagger}
The tagger takes global feature vectors as input and predicts a universal part of speech ([upos]{.smallcaps}), a language-specific tag ([xpos]{.smallcaps}), and morphological features ([ufeats]{.smallcaps}) for each word. The tagger consists of two linear layers followed by a softmax. Morphological features build a disordered set of category-value pairs (e.g. Number=Plur). Morphological feature prediction is thus implemented as several classification problems. The value of each morphological category is predicted with a FC network. Different parts of speech are assigned different sets of morphological categories (e.g. a noun can be attributed with grammatical gender, but not with grammatical tense). The set of possible values is thus extended with the NA (not applicable) symbol. It allows the model to learn that a particular category is not a property of a word.
#### Lemmatisation Module {#sec:lemmatizer}
The lemmatiser uses an approach similar to character-based word embedding estimation. A character embedding is concatenated with the global feature vector and transformed by a linear layer. The lemmatiser takes a sequence of such character representations and transforms it using a dilated CNN. The softmax function over the result produces the sequence of probabilities over a character vocabulary to form a lemma.
#### Parsing Module {#sec:parser}
Two single FC layers transform global feature vectors into head and dependent embeddings (see Figure [\[fig:tree\]](#fig:tree){reference-type="ref" reference="fig:tree"}). Based on these representations, a dependency graph is defined as an adjacency matrix with columns and rows corresponding to heads and dependents, respectively. The adjacency matrix elements are dot products of all pairs of the head and dependent embeddings (the dot product determines the certainty of the edge between two words). The softmax function applied to each row of the matrix predicts the adjacent head-dependent pairs. This approach, however, does not guarantee that the resulting adjacency matrix is a properly built dependency tree. The Chu-Liu-Edmonds algorithm is thus applied in the last prediction step. The procedure of predicting words' grammatical functions (aka dependency labels) is shown in Figure [\[fig:labels\]](#fig:labels){reference-type="ref" reference="fig:labels"}. A dependent and its head are represented as vectors by two single FC layers. The dependent embedding is concatenated with the weighted average of (hypothetical) head embeddings. The weights are the values from the corresponding row of the adjacency matrix, estimated by the arc prediction module. Concatenated vector representations are then fed to a FC layer with the softmax activation function to predict dependency labels.
#### EUD Parsing Module
Enhanced Universal Dependencies (EUD) are predicted similarly to dependency trees. The EUD parsing module is described in details in.
# COMBO Performance {#sec:evaluation}
#### Data
COMBO is evaluated on treebanks from the Universal Dependencies repository, preserving the original splits into training, validation, and test sets. The treebanks representing distinctive language types are summarised in Table [\[tab:statistics\]](#tab:statistics){reference-type="ref" reference="tab:statistics"} in Appendix [\[sec:data_statistics\]](#sec:data_statistics){reference-type="ref" reference="sec:data_statistics"}. By default, pre-trained 300-dimensional fastText embeddings are used. We also test encoding data with pre-trained contextual word embeddings (the tested BERT models are listed in Table [\[tab:berts\]](#tab:berts){reference-type="ref" reference="tab:berts"} in Appendix [\[sec:data_statistics\]](#sec:data_statistics){reference-type="ref" reference="sec:data_statistics"}). The UD datasets provide gold-standard tokenisation. If BERT intra-tokeniser splits a word into sub-words, the last layer embeddings are averaged to obtain a single vector representation of this word.
#### Qualitative Evaluation {#sec:resultOverview}
Table [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"} shows COMBO results of processing the selected UD treebanks.[^1] COMBO is compared with Stanza and spaCy.[^2] The systems are evaluated with the standard metrics: [F1]{.smallcaps}, [uas]{.smallcaps} (unlabelled attachment score), [las]{.smallcaps} (labelled attachment score), [mlas]{.smallcaps} (morphology-aware [las]{.smallcaps}) and [blex]{.smallcaps} (bi-lexical dependency score).[^3] COMBO and Stanza undeniably outrun spaCy models. COMBO using non-contextualised word embeddings is outperformed by Stanza in many language scenarios. However, COMBO supported with BERT-like word embeddings beats all other solutions and is currently the SOTA system for morphosyntactic analysis. Regarding lemmatisation, Stanza has an advantage over COMBO in most tested languages. This is probably due to the fact that Stanza lemmatiser is enhanced with a key-value dictionary, whilst COMBO lemmatiser is fully neural. It is not surprising that a dictionary helps in lemmatisation of isolating languages (English). However, the dictionary approach is also helpful for agglutinative languages (Finnish, Korean, Basque) and for Arabic, but not for Polish (fusional languages). Comparing COMBO models estimated with and without BERT embeddings, we note that BERT boost only slightly increases the quality of lemma prediction in the tested fusional and agglutinative languages.
For a complete insight into the prediction quality, we evaluate individual [upos]{.smallcaps} and [udeprel]{.smallcaps} predictions in English (the isolating language), Korean (agglutinative) and Polish (fusional). Result visualisations are in Appendix [\[sec:visualisations\]](#sec:visualisations){reference-type="ref" reference="sec:visualisations"}. COMBO took part in IWPT 2021 Shared Task on Parsing into Enhanced Universal Dependencies, where it ranked 4th.[^4] In addition to [elas]{.smallcaps} and [eulas]{.smallcaps} metrics, the third evaluation metric was [las]{.smallcaps}. COMBO ranked 2nd, achieving the average [las]{.smallcaps} of 87.84%. The score is even higher than the average [las]{.smallcaps} of 86.64% in Table [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"}, which is a kind of confirmation that our evaluation is representative, reliable, and fair.
#### Downstream Evaluation
According to the results in Table [\[tab:2\]](#tab:2){reference-type="ref" reference="tab:2"}, COMBO predicts high-quality dependency trees and parts of speech. We therefore conduct a preliminary evaluation of morphosyntactically informed word embeddings in the textual entailment task (aka natural language inference, NLI) in English and Polish. We compare the quality of entailment classifiers with two FC layers trained on max/mean-pooled BERT embeddings and sentence representations estimated by a network with two transformer layers which is given morphosyntactically informed word embeddings (i.e. BERT-based word embeddings concatenated with [upos]{.smallcaps} embeddings, [deprel]{.smallcaps} embeddings, and BERT-based embeddings of the head word). The morphosyntactically informed English NLI classifier achieves an accuracy of 78.84% and outperforms the max/mean-pooled classifiers by 20.77 pp and 5.44 pp, respectively. The Polish syntax-aware NLI classifier achieves an accuracy of 91.60% and outperforms the max/mean-pooled classifiers by 17.2 pp and 7.7 pp, respectively.
#### Efficiency Evaluation
We also compare spaCy, Stanza and COMBO in terms of their efficiency, i.e. training and prediction speed.[^5] According to the results (see Tables [\[tab:5\]](#tab:5){reference-type="ref" reference="tab:5"} and [\[tab:6\]](#tab:6){reference-type="ref" reference="tab:6"}), spaCy is the SOTA system, and the other two are not even close to its processing time. Considering COMBO and Stanza, whose prediction quality is significantly better than spaCy, COMBO is 1.5 times slower (2 times slower with BERT) than Stanza in predicting, but it is definitely faster in training. The reason for large discrepancies in training times is the different architecture of these two systems. Stanza is a pipeline-based system, i.e. its modules are trained one after the other. COMBO is an end-to-end system, i.e. its modules are jointly trained and the training process is therefore faster.
## Extra Facilities
COMBO includes two innovative modules for thematic role labelling and for sentence embedding estimation. Thematic label prediction is evaluated on an extended version of UD-like Polish dependency treebank ,[^6] in which the thematic roles of some dependents are in the 11th column. The quality of COMBO semantic labelling is 79.61 of SLAS F\(_1\)[^7] (precision of 79.67% and recall of 79.55%). Evaluation of COMBO sentences embeddings requires large-scale NLU research.[^8] It is outside the scope of this paper and we intend to conduct such research in the future.
# Getting Started with COMBO {#sec:getting_started}
#### Prediction
COMBO provides two main prediction modes: a Python library and a command-line interface (CLI). The Python package mode supports automated model download. The code snippet demonstrates downloading a pre-trained Polish model and processing a sentence:
To download a model for another language, select its name from the list of pre-trained models.[^9] The Python mode also supports acquisition of [deprel]{.smallcaps} or [upos]{.smallcaps} embeddings, for example:
In CLI mode, COMBO processes sentences using either a downloaded model or a model trained by yourself. CLI works on raw texts and on the CoNLL-U files (i.e. with tokenised sentences and even morphologically annotated tokens):
#### Model Training
COMBO CLI allows to train new models for any language. The only requirement is a training dataset in the CoNLL-U/CoNLL-X format. In the default setup, tokenised sentences are input and all possible predictors are trained:
If we only train a dependency parser, the default setup should be changed with configuration flags: `--features`{.bash} with a list of input features and `--targets`{.bash} with a list of prediction targets.
# Conclusion {#sec:conclusions}
We have presented COMBO, the SOTA system for morphosyntacic analysis, i.e. part-of-speech tagging, morphological analysis, lemmatisation, and (enhanced) dependency parsing. COMBO is a language-agnostic and format-independent system (i.e. it supports the CoNLL-U and CoNLL-X formats). Its implementation as a Python package allows effortless installation, and incorporation into any Python code or usage in the CLI mode. In the Python mode, COMBO supports automated download of pre-trained models for multiple languages and outputs not only categorical morphosyntactic features, but also their embeddings. In the CLI mode, pre-trained models can be manually downloaded or trained from scratch. The system training is fully configurable in respect of the range of input features and output predictions, and the method of encoding input data. Last but not least, COMBO maintains a balance between efficiency and quality. Admittedly, it is not as fast as spaCy, but it is much more efficient than Stanza considering the training time. Tested on the selected UD treebanks, COMBO morphosyntactic models enhanced with BERT embeddings outperform spaCy and Stanza models. | {'timestamp': '2021-09-14T02:16:40', 'yymm': '2109', 'arxiv_id': '2109.05361', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05361'} |
null | null |
# Introduction
SmB\(_6\) is a strongly correlated Kondo insulator that was recently found to have a robust surface state. It shows high--temperature metallic behavior which changes to an insulating one below 40 K as an energy gap \(\Delta \approx\) 15--20 meV opens. This energy gap, brought in by hybridization between the localized Sm 4*f* states and weakly correlated (mostly) the Sm 5*d* states, should lead to a diverging resistance at lower temperatures. Instead, the resistance flattens below 5 K, which imply a metallic surface state. All the transport features, including the activated behavior below approximately 40 K and the resistance plateau for \(T \lesssim\) 5 K, agree with the theoretical prediction that SmB\(_6\) is a correlated topological insulator (TI). In this description, the bulk of SmB\(_6\) is insulating, and non--trivial topological protected surface states are responsible for the low--temperature resistance saturation. Most of the recent experimental research on SmB\(_6\) shows conclusively that metallic surface states lie behind the leveling--off of its low--temperature resistivity. However, their origin and topological nature are still under debate.[@hlav] Furthermore, some results from specific heat, optical conductivity, and quantum oscillations measurements indicate a bulk origin of SmB\(_6\) metal--like behavior at low temperatures. There is currently no clear consensus on this point. The correlated state of SmB\(_6\) seems only slightly affected by disorder or light doping with non--magnetic impurities. The robustness of the bulk gap and the protection against disorder seem to be intrinsic to this material. All these features make SmB\(_6\) an attractive candidate for applications. Most of the research on this system has been focused on its equilibrium properties. However, samarium hexaboride shows interesting transient behavior which is not well studied. A large nonlinear resistance and an oscillating response upon biasing single crystals with a small *dc* current at low temperatures are two examples of such transient phenomena. A model which involves both surface and bulk states, in addition to a strong coupling between the thermal and the electrical properties, has recently been put forward to explain this response. Nevertheless, there are only few reports on the frequency--dependent transport properties of SmB\(_6\). Here, we report results from impedance spectroscopy (IS) experiments on SmB\(_6\), performed at low temperatures and over a wide frequency range. By applying a single frequency voltage to the sample, the phase shift and amplitude of the resultant current can be determined. Both the magnitude (resistive) and phase shift (reactive) component of a frequency--dependent impedance yield relevant information about intrinsic and extrinsic dielectric relaxation processes. In a linear approach, the total response of the system can be described by inductive, capacitive, and resistive elements of an equivalent electrical network. These components are uniquely defined and give a clear and simple description of the underlying physical processes involved. The remainder of the paper is as follows. The experimental procedure is described in Section II. Results of measurements are reported and discussed in Section III, and conclusions are drawn in Section IV.
# Experiment
Ac impedance measurements on stoichiometric and non--stoichiometric SmB\(_6\) single crystals were carried out in the 1.8 to 300 K temperature range. SmB\(_6\) crystals were grown by the floating zone technique using a high power xenon--arc--lamp--image furnace. This method yields large single crystals free of contamination although vacancies and defects might be still present. Samples used in the experiments were selected from a batch of crystals based on their quality and size. The surfaces of these crystals, (100 planes), were carefully etched, using a mixture of hydrochloric acid and water, to remove possible oxide layers. We used silver paint, gold or pure indium to prepare contacts for impedance measurements, on recently cleaved, about 200 \(\mu\)m thick samples. We did not observe any important features in our measurements which could be related to contacts we have applied. Impedance spectra were obtained with a precision impedance analyzer (Wayne-Kerr) in the 20 Hz up to 10 MHz frequency range. An excitation amplitude of 0.01V was applied to ensure the ohmic regime with a low signal--to--noise ratio. Before connecting the sample, open--circuit and short--circuit tests, in addition to high frequency compensation were carried out, to minimize parasitic contributions from the measurement setup. IS measurements were performed in both out--of plane and in--plane configurations shown in Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}. In the former configuration, a parallel plate capacitor configuration, stray capacitance may form on the edges of the electrodes. This causes a measurement error that can be avoided using the guard electrode which absorbs the electric field at the edge. We applied this guarded method in out--of plane studies (see Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}c). On the other hand, for an in--plane configuration, reducing the electrode spacing diminishes the proportion of the current carried through the bulk and increases the measurement sensitivity to the surface properties (see Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}a). To complete our study, current--voltage (\(I-V\)) characteristics were measured at low temperatures. The SmB\(_6\) samples can be driven into a dissipative regime with increasing current below about 5 K, where the \(I-V\) relation shows nonlinear negative differential resistance (NDR) from self heating. To properly measure temperature changes arising from current flow, we used a differential chromel vs. Au-0.07 at.% Fe thermocouple between the cryostat base and the top of the sample.
# Results and Discussion
Below we show and discuss the impedance spectra for several single crystalline SmB\(_6\) samples. To analyze experimental data we have to ensure that the observed variations of complex impedance do arise from an intrinsic mechanism and not from the extrinsic Maxwell--Wagner type polarization effects brought about by interfaces. Most interfaces, like a Schottky barrier at the electrode--sample contact or a thin insulating surface layer with a slightly different stoichiometry, may behave as a parallel-plate capacitor with a very large capacitance and, consequently, can give rise to a very high apparent value of the dielectric constant. To rule out these effects, we measured SmB\(_6\) single crystals using different contacts in various configurations. We contacted our samples with a silver conducting paste, sputtered gold, and by a direct deposition of pure In or In/Ga alloy. The variation of contact type did not lead to any significant change in impedance spectra. Therefore, we are quite confident that the reported spectra are intrinsic to SmB\(_6\). We give measured quantities in absolute values. Specific values, which are necessarily determined from the bulk sample geometry, are wrong in the surface--conduction dominated region as the thickness of topologically protected metallic surface states is not reliably known. Here, we discuss data obtained for \(T\lesssim\) 15 K where all interesting features happen. In the temperature range beyond 15 K, the total impedance signal, which is very small and quite noisy, corresponds to the bulk semiconductor only. In Fig. [\[S8-Z-T\]](#S8-Z-T){reference-type="ref" reference="S8-Z-T"}, we plot the real \(Re(Z)\) and imaginary \(Im(Z)\) part of impedance \(Z\) for a single crystalline SmB\(_6\) (sample S8) as a function of temperature \(T\), in the range from 2 to 20 K, for various frequencies. These data have been obtained for the in-plane configuration, with a probe separation of approximately 150 \(\mu\)m. (see Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}a)). Upon lowering the temperature, \(Re(Z)\) starts to increase quite sharply at about 20 K and levels out below \(T \lesssim\) 5 K. Previous studies of high quality SmB\(_6\) single crystals have shown the evolution of this system from a semi-metallic state at room temperature to a Kondo correlated insulating state, and finally to a very low carrier density metallic state at low temperatures. It is now generally accepted that, for \(T \lesssim\) 5 K, the transport in SmB\(_6\) takes place through surface states (SS), as the bulk resistance \(R_b\) is much higher than the surface resistance \(R_s\). With increasing temperature (4 \(\lesssim T \lesssim\) 10 K), \(R_b\) becomes comparable to that of the SSs, and conduction proceeds through both the bulk of the crystal and the surface channels. Beyond 10 K and up to approximately 40 K, the bulk resistance over this temperature range is much smaller than that of the surface states, and the latter have no effect on electrical transport. The data plotted in Fig. [\[S8-Z-T\]](#S8-Z-T){reference-type="ref" reference="S8-Z-T"} follow this scheme. Assuming that the surface conductivity is independent of temperature, we obtain for the bulk resistance a thermally activated behavior \(R_b \propto exp(\Delta E/k_B T)\), with an activation energy of 3.3 meV, for 2 K\(< T <\)`<!-- -->`{=html}10 K. The imaginary part of \(Z(T)\) for the sample S8 also shows a pronounced rise below 5 K for large excitation frequencies, followed by leveling out at lower temperatures (see inset in Fig. [\[S8-Z-T\]](#S8-Z-T){reference-type="ref" reference="S8-Z-T"}). The Kondo insulator SmB\(_6\) appears to have an intrinsic metal to high-dielectric material transition. As its insulating gap opens completely at low temperatures, the bulk becomes an insulator with a large dielectric constant, whereas the surface states are metallic. This can be represented by an equivalent parallel \(RC\) circuit and a topological surface state. Figures [\[S8\]](#S8){reference-type="ref" reference="S8"}(a) and (b) show the IS data plotted as the negative imaginary \(-Im(Z)\) vs. the real part \(Re(Z)\) (Cole--Cole plot) at various temperatures in the range from 2 to 10 K. In such impedance plots, one or more semicircles may be seen for a parallel resistor-capacitor \(RC\) circuit measurement; each semicircle corresponding to a dielectric relaxation arising from different, extrinsic (interfaces) or intrinsic (sample), contributions in series. Our data show one single, non-ideal dielectric relaxation, evidenced by one depressed semicircle for each temperature, where the center of each semicircle has shifted below the x-axis (\(Re(Z)\)). For a Debye--like relaxation (an ideal parallel resistor-capacitor element), a perfect semicircle is expected. Since the data cannot be modeled solely using standard \(RC\) elements, we used the alternative equivalent circuit depicted in Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}(b), which yields good fits at all frequencies. The "depressed\" behavior can be fitted by adding a series \(R_sL_s\) branch, in parallel to the conventional \(RC\) element. Physically, the \(R_bC_b\) part accounts for the insulating SmB\(_6\) bulk, with a standard dielectric behavior. (Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}(b)). The \(R_sL_s\) branch corresponds to the inductive response, most likely from the sample surface. We also examined the equivalent circuit, displayed in the inset of Fig. [\[S8\]](#S8){reference-type="ref" reference="S8"}(b), which gives a somewhat better fit for the sample S-8 data. The Constant Phase Element (\(CPE\)) has an impedance of the form \(Z(CPE) = 1/(Q_o(i\omega)^n)\) where \(Q_o\) models the effects of a non-ideal relaxation process or retardation. The exponent \(n\) is a measure of the non-ideality of the relaxation process, brought about by non-uniformity of physical properties of the system. When \(n\) = 1, the system is described by a single time-constant and the parameter \(Q_o\) has units of capacitance. Its variation with \(T\), obtained from fittings to the circuit in the inset of Fig. [\[S8\]](#S8){reference-type="ref" reference="S8"}(b), is shown in Fig. [\[S8\]](#S8){reference-type="ref" reference="S8"}(c). The fitting parameters we have extracted from our model for the S8 sample are shown in Figs. [\[S8\]](#S8){reference-type="ref" reference="S8"} (d)--(g), where each point on the curves represents a fit at a specific temperature. For this sample, \(R_s = 288.5 \Omega\). As discussed above, the bulk resistance \(R_b\) follows thermally activated behavior with an activation energy of 3.3 meV for 2 K\(< T <\)`<!-- -->`{=html}10 K. The equivalent sample resistance \(R_{eq}\) is calculated as \((1/R_b +1/Rs)^{-1}\). It is worth noticing that \(R_b\) and \(R_{eq}\) are very close above 6 K, suggesting that the current flows mainly through bulk beyond this temperature. Fig. [\[S8\]](#S8){reference-type="ref" reference="S8"}(g) depicts the sample capacitance \(C_b\) we have estimated from the model. The capacitance stays constant as the temperature increases up to about 6 K, where it upturns. For the in-plane configuration (see Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}(a)), the field lines of the applied *ac* electric field are not parallel in the sample and the changes in capacitance may not necessarily reflect intrinsic changes in the bulk dielectric permittivity. Figure [\[S8\]](#S8){reference-type="ref" reference="S8"}(f) shows temperature variation of fitted values for the inductance \(L_s\). In the low--temperature metallic state below 4K, \(L_s\) \(\approx\) 10\(^{-6}\) H. As the temperature increases, \(L_s\) drops by four orders of magnitude between 4 and 6 K, and is negligibly small farther on. We note that the temperature variations of \(L_s\) and \(R_{eq}\) are alike in the temperature range considered. The large drop of \(L_s\) occurs at the transition from surface to bulk dominated conduction. The inductive contribution in the metallic regime may arise from the coupling of 2D (metallic) and 3D (bulk) conduction channels, likely through defects.. On the other hand, recent scanning microwave impedance microscopy results show presence of one--dimensional conducting channels terminating at surface step edges. These conducting channels may contribute significantly to the total inductance at low temperatures. The IS data for SmB\(_6\) (sample S9) are reported in Fig. [\[S9-all\]](#S9-all){reference-type="ref" reference="S9-all"}. This sample is identical to the S8 sample, except for the contact separation (approximately 25 \(\mu\)m) which is almost one order of magnitude smaller than in sample S8. Therefore, we expect surface effects to be more pronounced. Indeed, the total impedance increase at high frequencies shows behavior typical of inductive contributions. The imaginary impedance \(Im(Z)\) is then positive at 1.8 K because the spectra are dominated by the surface inductance \(L_s\). The model applies just the same. The IS results obtained for the out-of plane configuration (see Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}(b)) of SmB\(_6\) single crystals (samples S1--S4) are much the same as the results for the in-plane configuration. Figure [\[S4\]](#S4){reference-type="ref" reference="S4"}(c) shows the Niquist plot for sample S4 at several temperatures in the range 2 K-10 K. Variations of the magnitude of the impedance \(Z(T)\) and of the phase angle \(\theta(T)\) at various frequencies are displayed in Figs. [\[S4\]](#S4){reference-type="ref" reference="S4"}(a) and (b). The values of the parameters, obtained from the fit to the equivalent circuit of Fig. [\[Conf\]](#Conf){reference-type="ref" reference="Conf"}, are plotted in Figs. [\[S4\]](#S4){reference-type="ref" reference="S4"}(d)--(f). Including the \(CPE\) element in the equivalent circuit instead of \(C_b\) does not improve our fittings. As discussed above, the inductive contribution is important below 4 K, drops by several orders of magnitude with increasing temperature and becomes negligible beyond 6 K. Our fits to experimental data were carried out with the least--square method. The relative error in parameters \(L_s\) and \(C_b\) is less than 5% for \(T \lesssim\) 5 K. However, for \(T \ge 6 K\), the fitting parameters may be more uncertain because of the small signal measured. In addition, we observe some parasitic stray inductance for the sample S8 (in--plane configuration) which bends curves in the Cole--Cole plot very slightly to the right for \(T \ge 6 K\). Since any contribution to the impedance from the experimental setup (wires plus apparatus) should be temperature independent, this is brought, most likely, by electrodes. In Fig. [\[I-V\]](#I-V){reference-type="ref" reference="I-V"}, we plot current-voltage (\(I-V\) and \(dV/dI\)) characteristics measured at low temperatures for sample S4. \(I-V\) relation shows nonlinear negative differential resistance (NDR) with increasing current below about 5 K. Figure [\[I-V\]](#I-V){reference-type="ref" reference="I-V"}(d) displays the change in sample temperature corresponding to this process. The observed increment in temperature, \(\Delta T\), is less than 0.6 K over the base temperature of 2 K and is brought about by self heating. Similar behavior, albeit with a larger \(\Delta T\), for SmB\(_6\) single crystals has been reported in Ref. . In addition, the authors of this report find large voltage oscillations driven by a small *dc* current. To account for this, a model, in which a coupling between metallic surface conduction and thermally activated bulk conduction giving rise to a large variation in sample heating, was proposed. Qualitatively, the Joule heating in the surface conduction regime rises locally the sample temperature. As a result, the sample resistance decreases, enabling bulk conduction, and thus reducing Joule self heating. Consequently, the sample temperature drops to the initial value and drives the system back to the surface conduction regime. This model, for a rather narrow range of the parameters, yields voltage oscillations that agree with experiment. In our set--up, we do not find self-sustained oscillations in the SmB\(_6\) crystals. The dynamics of the model reported in Ref. depends importantly on the ratio of the bulk resistance to the surface resistance in thermal equilibrium. This ratio is about or less than 80 in the crystals of Refs. , and , whereas it is larger than \(10^4\) for the SmB\(_6\) samples we have studied. As a result, raising the sample temperature by 1 K in the former case (from a base temperature of 2 K) is sufficient to drive SmB\(_6\) crystals into a bulk conduction regime but our samples must be heated at least 5 K in order to reach the same conduction regime. On this account, the sample heating upon biasing with a *dc* current is much smaller in our experiments. In addition, the model of Ref. involves local effects and it is not known which portion of the sample is affected by current. It is not so in the experimental configuration we have used in which the whole sample is subjected to heating. These features might have hindered the non-linear oscillations found in Ref. as well as the shape (surface area and thickness) of the SmB\(_6\) platelets. On the other hand, it is well known that many dissipative systems, such as inductor-capacitor-resistor (\(LCR\)) circuits with a nonlinear negative--differential resistance show self-sustained oscillations which can be described by the van der Pol equation. As shown above, the SmB\(_6\) low-temperature impedance spectra can be interpreted by an equivalent \(LCR\) circuit, with NDR coming from the self heating. It is therefore reasonable to ask whether the self-sustained voltage oscillations in SmB\(_6\) are the ones depicked by the van der Pol equation. If so, the oscillation frequency should be inversely proportional to the square root of the effective capacitance times inductance, i.e., \(LC \propto 1/\omega^2\). This verifies that SmB\(_6\) crystals may work as a current-controlled oscillator in the 20 MHz frequency range as shown in Ref. .
# Conclusions
In-plane and out-of-plane impedance spectra were measured in SmB\(_6\) single crystals in order to study the variation of the dielectric response of this system at low temperatures. A universal model describes quantitatively the complex impedance data in surface-and bulk--conduction dominated regimes. The model includes a \(R_bC_b\) circuit and an inductive \(R_sL_s\) branch in parallel. The latter becomes dominant at high frequencies. The \(R_bC_b\) element is made of a resistor and a capacitor connected in parallel. It describes standard dielectric relaxations in bulk. The \(R_sL_s\) branch accounts for the inductive contribution seen below 5 K, in the surface--conduction metallic regime. The equivalent inductance, obtained form fits to experimental data, drops drastically with increasing temperature as the bulk starts to control electrical conduction. We attribute this inductive contribution to the surface states and its coupling to the bulk at low temperatures. Platelet shaped SmB\(_6\) single crystals show a current--controlled negative--differential resistance at low temperatures, which is brought by a Joule heating. The inductive and capacitive contributions to the impedance, obtained for SmB\(_6\) single crystals at low temperatures, can give rise to the self--sustained voltage oscillations found in Ref. . We acknowledge the financial support of Ministerio de Ciencia, Innovación y Universidad of Spain through Grant RTI2018-098537-B-C22 co--funded by ERDF, EU. The work at the University of Warwick was supported by EPSRC, UK, through Grant EP/T005963/1. | {'timestamp': '2021-09-14T02:14:23', 'yymm': '2109', 'arxiv_id': '2109.05313', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05313'} |
# PRELIMINARIES AND RESULTS {#preliminaries-and-results .unnumbered}
The idea of category base is a generalization of both measure and topology and its main objective is to present measure and Baire category(topology) and also some other aspects of point set classification within a common framework. It was introduced by J.C.Morgan II in the mid seventies of the last century and has developed since then through a series of papers \([1],[2],[3],[4],[5],\) etc. To start with, we recall some basic definitions and theorems which may be found in the above references and also in the monograph \[6\].
Several examples of category bases can be found in \[6\].
If countable sets are meager, then the category base is point-meager base and a category base whose regions are abundant sets is a Baire base.
In \[6\], one can find several examples corresponding to Definition 2 and Definition 3.\
From Definition 3, it follows that a set is non-Baire if there is a region (which is not necessarily unique) in every subregion of which both the set and its complement are abundant. In this paper, we identify this special region as a test region corresponding to the non-Baire set.\
Below we establish two alternative characterizations for non-Baire sets in a category base. We define
The above definition is somewhat analogous to the notion of a completely I-nonmeasurable set, given in \[8\].
With reference to this definition, we prove the following decomposition theorem.
Evidently, any \(\mathcal{M}\)-family is a special case of a category base, so the definitions of meager, abundant and Baire sets are as usual. Moreover, any \(\mathcal{M}\)-family is point-meager and every member of it is an abundant set which follows directly from the axioms (1) and (2) in the above Definition.\
If \(\mathcal{C}\) is a non-empty family of subsets of a non-empty set X and S is a subset of X, then as described in \[3\], the generalized Banach-Mazur game \(\Gamma\)(S, \(\mathcal{C}\)) is played as follows: Two players I and II, alternatively choose sets from \(\mathcal{C}\) to define a descending sequence of sets, player I selecting the sets in the sequence with odd indices and player II selecting the sets with even indices. If the intersection of the constructed sequence has atleast one point (which obviously exists if \(\mathcal{C}\) is an \(\mathcal{M}\)-family) in S, then player I wins; otherwise player II wins. In the game \(\Gamma\)(S, \(\mathcal{C}\)), let us denote player I, player II by using symbols \(\langle\)S\(\rangle\), \(\langle\)X\(-\)S\(\rangle\) respectively.\
Now suppose C\(\in\]\mathcal{C}\) such that S\(\cap\)C \(\neq\) \(\phi\) \(\neq\) (X\(-\)S)\(\cap\)C. Then we can describe the game \(\Gamma\)(S,\(\mathcal{C}_{|C}\)) in a manner similar as above where \(\mathcal{C}_{|C}\)={E\(\in\)\(\mathcal{C}\): E\(\subseteq\)C}. As a consequence of the Theorem 2 \[3\], Theorem 11 may be interpreted in terms of the generalized Banach-Mazur game in the following manner : | {'timestamp': '2021-09-14T02:22:03', 'yymm': '2109', 'arxiv_id': '2109.05499', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05499'} |
null | null |
null | null |
null | null |
# Introduction {#intro}
In recent years, we have witnessed a burst of quaternion representations in many fields, including color imaging, signal processing, robotics, rolling bearing fault diagnosis, quaternion convolutional neural networks (QCNNs), etc. Moreover, there are some noticeable steps towards optimizing the corresponding quaternion represented problems. Specifically, Qi et al. conducted a systematic study on quaternion matrix optimization, and Flamant et al. proposed a general framework for constrained convex quaternion optimization. In terms of algorithms in the quaternion domain, affine projection algorithms and learning algorithms based on gradient and Hessian have been proposed and analyzed. Hence, the increasing number of quaternion-represented applications and the studies on the associated optimization problems call for a deeper understanding of the quaternion structure that could lead to some efficient solution methods. In this paper, we shall focus on one algebraic quaternion structure: matrix rank-one decomposition, and show that such decomposition admits a stronger property than that in the real and complex domain by leveraging the intrinsic quaternion nature. We further show that such merit of rank-one decomposition can be extended to some of its theoretical implications such as \(\mathcal{S}\)-Procedure, joint numerical range and quadratically constrained quadratic programming (QCQP) in the quaternion domain, and improve the associated results in the real and complex domains. The matrix rank-one decomposition that we discuss in this paper is a technique of decomposing a positive semidefinite Hermitian matrix into the sum of rank-one matrices to satisfy the so-called equal inner product property, i.e., the inner product between some given matrices and each rank-one term in the decomposition has the same value. The first such type of decomposition was introduced by Sturm and Zhang in with the equal inner product property valid for *one* matrix, and it was used as a key technique to establish the Linear Matrix Inequality presentation of a class of matrix cones with its quadratic form co-positive over the real domain. Moreover, such a decomposition technique was found to be useful in quadratic minimization and designing approximation algorithms for biquadratic optimization. Soon after the work of, Huang and Zhang extended the matrix rank-one decomposition to the complex domain such that the equal inner product property holds for *two* matrices. Interestingly, we find that such property remains valid for *four* matrices when the rank-one decomposition is conducted in the quaternion domain. Moreover, our proof is fully constructive and the corresponding computational procedure is summarized in Algorithm [\[ACQMD1\]](#ACQMD1){reference-type="ref" reference="ACQMD1"}. Theoretical implications of our novel matrix decomposition technique are quite versatile and yield stronger results than those in the real and complex domains. The first two theoretical implications are in joint numerical range and \(\mathcal{S}\)-Procedure, both of which have some fundamental impacts and wide applications in many fields. In particular, the joint numerical range is an important tool in linear algebra and convex analysis, and it is found to be useful in spectrum analysis and quantum computing. \(\mathcal{S}\)-Procedure occupies a crucial position in the field of robust optimization, statistics, signal processing, among others. With our matrix rank-one decomposition result, we manage to establish the convexity of joint numerical range for *five* matrices and the *lossless* of \(\mathcal{S}\)-Procedure for *four* Hermitian forms. As a comparison, similar results only hold for fewer matrices in the real and complex domains. In addition, our result can also be applied to quadratically constrained quadratic programming (QCQP). To be specific, when the number of quadratic constraints is no larger than \(4\), we show that a rank-one solution of the SDP relaxation of (QCQP), which is hence an optimal solution of (QCQP), can be recovered from our matrix rank-one decomposition technique. This paper is organized as follows. In Section [2](#sec:1){reference-type="ref" reference="sec:1"}, we introduce some notations and definitions used throughout this paper. Section [3](#sec:qua_matrix_rank_one){reference-type="ref" reference="sec:qua_matrix_rank_one"} is devoted to the new quaternion matrix rank-one decomposition theorem. To showcase the capability of our new theorem, we illustrate some improved results in the joint numerical range in Section [4](#sec:joint_num_range){reference-type="ref" reference="sec:joint_num_range"} and the \(\mathcal{S}\)-Procedure in Section [5](#sec:s_lem){reference-type="ref" reference="sec:s_lem"}. Finally, we present how to solve the (QCQP) with our novel decomposition technique as another theoretical implication of our result in Section [6](#sec:qua_quadratic_opt){reference-type="ref" reference="sec:qua_quadratic_opt"}.
# Preliminaries {#sec:1}
In this section, we introduce some basic algebraic operations in the quaternion domain for scalars, vectors and matrices.
## Quaternion Operations for Scalars {#sec:qua_alg}
We define the set of quaternions \(\mathbb{H}\) as a 4-dimensional normed division algebra over the real numbers \(\mathbb{R}\). It has a canonical basis \(\{1, \mbox{\boldmath{\)i\(}}, \mbox{\boldmath{\)j\(}}, \mbox{\boldmath{\)k\(}}\}\), where \(\mbox{\boldmath{\)i\(}}, \mbox{\boldmath{\)j\(}}, \mbox{\boldmath{\)k\(}}\) are imaginary units such that \[\label{ima_multi} \mbox{\boldmath{\(i\)}}^2=\mbox{\boldmath{\(j\)}}^2=\mbox{\boldmath{\(k\)}}^2=\mbox{\boldmath{\(i\)}}\mbox{\boldmath{\(j\)}}\mbox{\boldmath{\(k\)}}=-1,\quad \mbox{\boldmath{\(i\)}}\mbox{\boldmath{\(j\)}}=-\mbox{\boldmath{\(j\)}}\mbox{\boldmath{\(i\)}}=\mbox{\boldmath{\(k\)}}.\] Then, any quaternion \(q \in \mathbb{H}\) can be written as \[q = q_a + q_b\mbox{\boldmath{\(i\)}} + q_c\mbox{\boldmath{\(j\)}} + q_d\mbox{\boldmath{\(k\)}},\] where \(q_a,q_b,q_c,q_d \in \mathbb{R}\) are the components of \(q\). The real and imaginary parts of \(q\) are denoted as \(\mbox{\rm Re~} q=q_a\) and \(\mbox{\rm Im~} q=q_b\mbox{\boldmath{\)i\(}} + q_c\mbox{\boldmath{\)j\(}} + q_d\mbox{\boldmath{\)k\(}}\) respectively. Note that, in contrast with the product operation in the real and complex domain, the product operation in the quaternion domain is noncommutative, i.e., \(q \cdot p \ne p \cdot q\) for \(p, q \in \mathbb{H}\). We denote by \(\overline{q}=\mbox{\rm Re~} q-\mbox{\rm Im~} q\) the quaternion conjugate of \(q\), and it holds that \(\overline{(p \cdot q)}=\overline{q} \cdot \overline{p}\). For a given quaternion \(q \in \mathbb{H}\), \(|q|\) denotes its modulus and can be expressed as \[|q|=\sqrt{q \cdot \overline{q}}=\sqrt{\overline{q} \cdot q}=\sqrt{q_a^2+q_b^2+q_c^2+q_d^2}.\] Then, we denote \(q = |q|(\cos\theta + q^I\sin\theta)\) as the triangle representation of quaternion \(q\), where \(\cos\theta = \frac{q_a}{|q|}\), \(\sin\theta = \frac{\sqrt{q_b^2 + q_c^2 + q_d^2}}{|q|}\) and \(q^I = \frac{q_b\mbox{\boldmath{\)i\(}} + q_c\mbox{\boldmath{\)j\(}} + q_d\mbox{\boldmath{\)k\(}}}{\sqrt{q_b^2 + q_c^2 + q_d^2}}\). At last, any non-zero quaternion \(q\) has an inverse \(q^{-1}=\overline{q}/|q|^2\) and the inverse of the product of two quaternions is \((p \cdot q)^{-1}=q^{-1} \cdot p^{-1}\).
## Quaternion Vectors and Quaternion Matrices
Similar to the scalar case, a quaternion vector \(\mbox{\boldmath{\)q\(}} \in \mathbb{H}^n\) can be written as \[\mbox{\boldmath{\(q\)}} = \mbox{\boldmath{\(q\)}}_a + \mbox{\boldmath{\(q\)}}_b\mbox{\boldmath{\(i\)}} + \mbox{\boldmath{\(q\)}}_c\mbox{\boldmath{\(j\)}} + \mbox{\boldmath{\(q\)}}_d\mbox{\boldmath{\(k\)}},\] where \(\mbox{\boldmath{\)q\(}}_a,\mbox{\boldmath{\)q\(}}_b,\mbox{\boldmath{\)q\(}}_c,\mbox{\boldmath{\)q\(}}_d \in \mathbb{R}^n\) are the components of \(\mbox{\boldmath{\)q\(}}\). For a quaternion vector \(\mbox{\boldmath{\)q\(}}\), \(\mbox{\boldmath{\)q\(}}^{\top}\) denotes the transpose of \(\mbox{\boldmath{\)q\(}}\), and \(\mbox{\boldmath{\)q\(}}^{H}=(\overline{\mbox{\boldmath{\)q\(}}})^{\top} = \overline{(\mbox{\boldmath{\)q\(}}^{\top})}\) denotes its conjugate transpose. Any quaternion matrix \(A\in\mathbb{H}^{m\times n}\) can be expressed as \[A = A_a+A_b\mbox{\boldmath{\(i\)}}+A_c\mbox{\boldmath{\(j\)}}+A_d\mbox{\boldmath{\(k\)}}\] with \(A_a,A_b,A_c,A_d \in \mathbb{R}^{m\times n}\). The transpose and the conjugate transpose of \(A\) are \(A^{\top}\) and \(A^H=(\overline{A})^{\top} = \overline{(A^{\top})}\), respectively. For two quaternion matrices \(A \in \mathbb{H}^{m \times n}\) and \(B \in \mathbb{H}^{n \times p}\), we have \((AB)^H = B^HA^H\); however, due to the noncommutativity of the product operation, \((AB)^\top \not= B^\top A^\top\) and \(\overline{(AB)} \not= \overline{A} \ \overline{B}\). For two vectors \(\mbox{\boldmath{\)q\(}}, \mbox{\boldmath{\)p\(}} \in \mathbb{H}^n\), their inner product \[\mbox{\boldmath{\(q\)}} \bullet \mbox{\boldmath{\(p\)}} = \mbox{\rm Re~}(\mbox{\boldmath{\(q\)}}^H\mbox{\boldmath{\(p\)}}) = \mbox{\boldmath{\(q\)}}_a^\top\mbox{\boldmath{\(p\)}}_a+\mbox{\boldmath{\(q\)}}_b^\top\mbox{\boldmath{\(p\)}}_b+\mbox{\boldmath{\(q\)}}_c^\top\mbox{\boldmath{\(p\)}}_c+\mbox{\boldmath{\(q\)}}_d^\top\mbox{\boldmath{\(p\)}}_d.\] Similarly, for two matrices \(A ,B \in \mathbb{H}^{n \times n}\), their inner product is defined as \[A \bullet B := \mbox{\rm Re~}(\mbox{\rm tr}\, A^HB)=\mbox{\rm tr}\, (A_a^\top B_a + A_b^\top B_b + A_c^\top B_c + A_d^\top B_d),\] where 'tr' denotes the trace of a matrix.
## Hermitian and Positive Semidefinite Matrices
We call \(X\) a quaternion Hermitian matrix if it satisfies \(X = X^H\). The cone of quaternion Hermitian matrices is denoted as \(\mathcal{H}^{n}\). Then \(\mbox{\boldmath{\)u\(}}^HX\mbox{\boldmath{\)u\(}}\) is real for all \(\mbox{\boldmath{\)u\(}} \in \mathbb{H}^n\) if and only if \(X \in \mathcal{H}^{n}\). We denote by \(\mathcal{S}_{+}^{n}(\mathcal{S}_{++}^{n})\), \(\mathcal{C}_{+}^{n}(\mathcal{C}_{++}^{n})\) and \(\mathcal{H}_{+}^{n}(\mathcal{H}_{++}^{n})\) the cones of real symmetric positive semidefinite (positive definite), complex Hermitian positive semidefinite (positive definite) and quaternion Hermitian positive semidefinite (positive definite) matrices, respectively. The notation \(X \succeq 0\) (\(X \succ 0\)) means that \(X\) is positive semidefinite (positive definite). Then, for any matrix \(X \succeq 0\) (\(X \succ 0\)), we have \(\mbox{\boldmath{\)u\(}}^HX\mbox{\boldmath{\)u\(}}\) is real and nonnegative (positive) for all \(\mbox{\boldmath{\)u\(}} \in \mathbb{H}^n\) (\(0 \neq \mbox{\boldmath{\)u\(}} \in \mathbb{H}^n\)).
# The Quaternion Matrix Rank-One Decomposition Method {#sec:qua_matrix_rank_one}
In this section, we discuss a particular rank-one decomposition of matrices over the quaternion domain such that the inner product between some given matrices and each rank-one term in the decomposition has the same value. In the real domain, Sturm and Zhang showed that for a rank-\(r\) symmetric positive semidefinite matrix \(Y \in \mathcal{S}_+^{n}\) and a real symmetric matrix \(B \in \mathcal{S}^{n}\), there is a rank-one decomposition of \(Y\) such that: \[Y = \sum_{i=1}^r\mbox{\boldmath{\(y\)}}_i\mbox{\boldmath{\(y\)}}_i^\top~~\mbox{\rm and}~~\mbox{\boldmath{\(y\)}}_j^\top B\mbox{\boldmath{\(y\)}}_j= (\mbox{\boldmath{\(y\)}}_i\mbox{\boldmath{\(y\)}}_i^\top) \bullet Y = \frac{B \bullet Y}{r},~~\mbox{for}~j=1,2,\cdots,r.\] Subsequently, Huang and Zhang proved that the above decomposition could hold for two matrices in the complex domain. In particular, suppose \(Z\in\mathcal{C}_+^n\) is a complex Hermitian positive semidefinite matrix of rank \(r\), and \(C_1,C_2 \in \mathcal{C}^n\) are two given complex Hermitian matrices. Then, there is a rank-one decomposition of \(Z\) such that: \[Z = \sum_{i=1}^r\mbox{\boldmath{\(z\)}}_i\mbox{\boldmath{\(z\)}}_i^H~~\mbox{\rm and}~~\mbox{\boldmath{\(z\)}}_j^H C_k\mbox{\boldmath{\(z\)}}_j = \frac{C_k \bullet Z}{r}, ~~\mbox{for}~j=1,2,\cdots,r~\mbox{\rm and}~k=1,2.\] Thus, a natural question arises: can a similar decomposition holds for more matrices over the quaternion domain? The following theorem is the main result of this paper and gives an affirmative answer to this question.
Note that the rank-one decomposition procedure we provide in the proof of Theorem [\[qua_decomposition\]](#qua_decomposition){reference-type="ref" reference="qua_decomposition"} is actually implementable. To make it clear, we summarize all the computational procedures in Algorithm [\[ACQMD1\]](#ACQMD1){reference-type="ref" reference="ACQMD1"} below. In particular, Step \(1\) and Step \(2\) in the proof are described by Algorithm [\[ACQMD1\]](#ACQMD1){reference-type="ref" reference="ACQMD1"} with \(\ell =3\) and \(\ell =4\), respectively.
With this concept in hand, we formally define the joint numerical range of a set of matrices as follows.
Regarding the convexity of the joint numerical ranges, Hausdorff showed that
In a slightly different form, Brickman extended the above result to three matrices, i.e., suppose \(A_1, A_2, A_3\) are \(n \times n\) complex Hermitian matrices, then \[\label{form-Brickman}\left\{ \begin{pmatrix} \mbox{\boldmath{\(x\)}}^HA_1\mbox{\boldmath{\(x\)}} \\ \mbox{\boldmath{\(x\)}}^HA_2\mbox{\boldmath{\(x\)}} \\ \mbox{\boldmath{\(x\)}}^HA_3\mbox{\boldmath{\(x\)}} \end{pmatrix} ~:~ \mbox{\boldmath{\(x\)}} \in \mathbb{C}^n \right\}\] is a convex set, where \(\mathbb{C}^n\) is the set of complex vectors. As a matter of fact, the convexity of the joint numerical ranges in complex domain can also be extended to quaternion domain for more matrices. In particular, Au-Yeung and Poon established the following result for quaternion Hermitian matrices.
Note that the above theorem assumes \(n \not= 2\) and this condition is necessary as a counterexample is presented in. However, such condition can be relaxed if we consider the slightly different form of [\[form-Brickman\]](#form-Brickman){reference-type="eqref" reference="form-Brickman"} by Brickman, where the restriction \(\mbox{\boldmath{\)x\(}}^H\mbox{\boldmath{\)x\(}}=1\) is dropped. We present this result in Theorem [\[joint_num_range_quaternion\]](#joint_num_range_quaternion){reference-type="ref" reference="joint_num_range_quaternion"}, which is proved by our quaternion matrix rank-one decomposition theorem.
# S-Procedure {#sec:s_lem}
It is well known that \(\mathcal{S}\)-procedure plays an important role in robust optimization, statistics, signal processing, control and stability problems, among others. In this section, we discuss another interesting theoretical implication of our new rank-one decomposition (Theorem [\[qua_decomposition\]](#qua_decomposition){reference-type="ref" reference="qua_decomposition"}) in \(\mathcal{S}\)-procedure. We first recall the following lemma, which is due to, about \(\mathcal{S}\)-procedure in real domain.
We remark that the above result essentially study the relationship between \[\label{S-procedure1} G_1(\mbox{\boldmath{\(x\)}}) \ge 0, G_2(\mbox{\boldmath{\(x\)}}) \ge 0, \cdots, G_m(\mbox{\boldmath{\(x\)}}) \ge 0 \Rightarrow F(\mbox{\boldmath{\(x\)}}) \ge 0, \vspace{-1.5mm}\] and \[\label{S-procedure2} \exists \tau_1 \ge 0, \tau_2 \ge 0, \cdots, \tau_m \ge 0 \ {\rm such \ that} \ F(\mbox{\boldmath{\(x\)}})-\sum_{i=1}^m \tau_iG_i(\mbox{\boldmath{\(x\)}}) \ge 0 \ \forall \mbox{\boldmath{\(x\)}}.\] It is obvious that [\[S-procedure2\]](#S-procedure2){reference-type="eqref" reference="S-procedure2"} implies [\[S-procedure1\]](#S-procedure1){reference-type="eqref" reference="S-procedure1"}, and the converse also holds when \(m=1\). In the following, we call \(\mathcal{S}\)-procedure is *lossless* if [\[S-procedure1\]](#S-procedure1){reference-type="eqref" reference="S-procedure1"} and [\[S-procedure2\]](#S-procedure2){reference-type="eqref" reference="S-procedure2"} are equivalent. Although \(\mathcal{S}\)-procedure was first studied in the real domain, interestingly, a stronger result has been established in the complex domain.
Inspired by such result, we further consider \(\mathcal{S}\)-procedure in the quaternion domain and manage to show that it is *lossless* for \(m = 4\) by Theorem [\[qua_decomposition\]](#qua_decomposition){reference-type="ref" reference="qua_decomposition"}.
# Quaternion Quadratically Constrained Quadratic Optimization {#sec:qua_quadratic_opt}
In this section, we consider the following quadratically constrained quadratic programming in the quaternion domain: \[\label{QCQP_stand} \begin{aligned} {\rm (QCQP)} \quad \max \quad & \mbox{\boldmath{\(x\)}}^HQ\mbox{\boldmath{\(x\)}} + 2{\rm Re}(\mbox{\boldmath{\(x\)}}^Hq)\\ {\rm s.t.} \quad & \mbox{\boldmath{\(x\)}}^HA_j\mbox{\boldmath{\(x\)}} + 2{\rm Re}(\mbox{\boldmath{\(x\)}}^Hb_j) + c_j \le 0, \ j = 1,\cdots,m, \end{aligned}\] where \(Q, A_j \in \mathcal{H}^n\), \(b_j, q \in \mathbb{H}^n\), \(c_j \in \mathbb{R}\), \(j = 1,\cdots,m\). Normally, solving the above problem is very challenging. However, when the number of constraints \(m\) is small, the problem is possibly tractable. For instance, in the complex domain, Huang and Zhang showed that the problem (QCQP) with \(m=2\) could be cast as an SDP and thus could be solved in polynomial time. While regarding the quaternion domain, we shall show that a similar result holds for a larger value of \(m\) (in particular for \(m=4\)) by the afore-mentioned rank-one decomposition in Theorem [\[qua_decomposition\]](#qua_decomposition){reference-type="ref" reference="qua_decomposition"}. To relate this problem to SDP, we rewrite (QCQP) as the following matrix form: \[\label{QCQP2} \begin{aligned} \quad \max \quad & B_0 \bullet \begin{bmatrix} 1 & \mbox{\boldmath{\(x\)}}^H \\ \mbox{\boldmath{\(x\)}} & \mbox{\boldmath{\(x\)}}\bx^H \\ \end{bmatrix}\\ {\rm s.t.} \quad & B_j \bullet \begin{bmatrix} 1 & \mbox{\boldmath{\(x\)}}^H \\ \mbox{\boldmath{\(x\)}} & \mbox{\boldmath{\(x\)}}\bx^H \\ \end{bmatrix} \le 0, ~j = 1,\cdots,m, \end{aligned}\] where \(B_0 = \begin{bmatrix} 0 & q^H \\ q & Q \\ \end{bmatrix}, \ B_j = \begin{bmatrix} c_j & b_j^H \\ b_j & A_j \\ \end{bmatrix}, ~j = 1,\cdots,m\). Problem [\[QCQP2\]](#QCQP2){reference-type="eqref" reference="QCQP2"} can be homogenized by introducing a new variable \(t\) and requiring \(|t|^2=1\): \[\label{HQ} \begin{aligned} {\rm (HQCQP)} \quad \max \quad & B_0 \bullet \begin{bmatrix} |t|^2 & t \cdot \mbox{\boldmath{\(x\)}}^H \\ \mbox{\boldmath{\(x\)}} \cdot \bar{t} & \mbox{\boldmath{\(x\)}}\bx^H \\ \end{bmatrix}\\ {\rm s.t.} \quad & B_j \bullet \begin{bmatrix} |t|^2 & t \cdot \mbox{\boldmath{\(x\)}}^H \\ \mbox{\boldmath{\(x\)}} \cdot \bar{t} & \mbox{\boldmath{\(x\)}}\bx^H \\ \end{bmatrix} \le 0, ~j = 1,\cdots,m, \\ & B_{m+1} \bullet \begin{bmatrix} |t|^2 & t \cdot \mbox{\boldmath{\(x\)}}^H \\ \mbox{\boldmath{\(x\)}} \cdot \bar{t} & \mbox{\boldmath{\(x\)}}\bx^H \\ \end{bmatrix} = 1, \end{aligned}\] where \(B_{m+1} = \begin{bmatrix} 1 & 0 \\ 0 & 0 \\ \end{bmatrix}\). Then for any solution \(\begin{pmatrix} t \\ \mbox{\boldmath{\)x\(}} \\ \end{pmatrix}\) of (HQCQP), since \(|t|^2=1\), \(\mbox{\boldmath{\)x\(}} \cdot \bar {t}\) is a solution to [\[QCQP2\]](#QCQP2){reference-type="eqref" reference="QCQP2"} and hence a solution to (QCQP). Therefore, we shall work on (HQCQP) in the rest of this section. By letting \[X = \begin{bmatrix} |t|^2 & t \cdot \mbox{\boldmath{\(x\)}}^H \\ \mbox{\boldmath{\(x\)}} \cdot \bar{t} & \mbox{\boldmath{\(x\)}}\bx^H \\ \end{bmatrix}~\mbox{and dropping the constraint:}~\mbox{\rm rank}\, (X) =1,\] we obtain an SDP relaxation of (HQCQP): \[\label{QCQPR} \begin{aligned} {\rm (QCQPR)} \quad \max \quad & B_0 \bullet X \\ {\rm s.t.} \quad & B_j \bullet X \le 0, ~j = 1,\cdots,m, \\ & B_{m+1} \bullet X = 1, \\ & X \succeq 0, \end{aligned}\] whose dual is given by \[\label{DQCQPR} \begin{aligned} {\rm (DQCQPR)} \quad \min \quad & y_0 \\ {\rm s.t.} \quad & Y = \sum_{j=1}^m y_jB_j-B_0 + y_0B_{m+1} \succeq 0, \\ & y_j \ge 0,~j = 1,\cdots,m,~y_0 \ {\rm free}. \end{aligned}\] To proceed, we assume (QCQP) satisfies the Slater condition, that is, there exists \(\mbox{\boldmath{\)x\(}}_0 \in \mathbb{H}^n\) such that \[\label{slater} q_j(\mbox{\boldmath{\(x\)}}_0) := B_j \bullet \begin{bmatrix} 1 & \mbox{\boldmath{\(x\)}}_0^H \\ \mbox{\boldmath{\(x\)}}_0 & \mbox{\boldmath{\(x\)}}_0\mbox{\boldmath{\(x\)}}_0^H \\ \end{bmatrix} < 0, j = 1,\cdots,m.\] Accordingly, (QCQPR) satisfies the Slater condition as well. Now we are ready to present the main theorem of this section, which states that (QCQP) is essentially equivalent to an SDP when \(m = 4\). | {'timestamp': '2021-09-14T02:18:16', 'yymm': '2109', 'arxiv_id': '2109.05405', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05405'} |
null | null |
null | null |
null | null |
# Introduction {#sec:introduction}
Learning (FL) is transforming many industries including defense, telecommunications, IoT, and transportation. This Machine Learning (ML) paradigm addresses issues such as data privacy, security, access rights and access to heterogeneous data by training a global model using distributed data. In FL, a group of nodes called *workers* participate to build a global model under the coordination of one node, the *chief*. After each round of the FL process, the *chief* sends a matrix of weights to the *workers*. Each *worker* then computes its local model update based on their local data. Once the *chief* node gets all the local model updates from the *workers*, it averages the models with an aggregation rule and updates them into a single matrix. The process repeats for several rounds of training until a pre-specified test accuracy is reached or the maximum number of time epochs have elapsed. Although FL enables nodes to construct a model without sharing their private data with others, the fact that the *chief* has no visibility into how the updates are generated represents a vulnerability in the process. Regarding the cybersecurity of FL, attacks aim at exploiting this vulnerability. According to Blanchard et al., classical averaging rules are so fragile that even a single malicious *worker* can manipulate the entire learning process by sabotaging prediction accuracy. Many defense mechanisms propose corrections to the FL process to increase system efficiency. They apply statistical methods to correct the aggregation rule implemented at the *chief* . Rather than correcting, other approaches proposed to detect malicious *workers* by interpreting information of the *workers*' behaviors in order to eliminate potential malicious *workers*. Kang et al. proposed a reputation-based scheme in order to select reliable and trusted *workers* . Preuveneers et al. proposed to audit the local model updates and the *workers* in the FL process are held accountable. As most of the defense strategies are based on malicious *worker* detection, there is an urgency to prepare the federated learning process for monitoring and eventually behavioral pattern analysis. In fact, without timely and dynamic monitoring methods, the *chief* cannot detect and remove the malicious or unreliable *workers* from the system. Moreover, it is hard for the *chief* to monitor the large-scale *worker* behaviors in real-time. Pan et al. scratched the surface by proposing to visualize the behavior of the *workers*. They suggested to visualize the policy sequence of their model in order to understand the temporal patterns of the attacks. This paper proposes to study the information inside the learning process of each *worker* in order to increase the efficiency of the behavior pattern analysis approaches. Precisely, we aim to analyse at what time is it appropriate to start the detection algorithm in order to remove the malicious or unreliable *workers* from the system. The intuition is that *workers* should not be removed from the FL process too early nor too late. There must be an appropriate time window for monitoring after which the detection protocol or the behavior pattern analysis will start declaring that the nodes are malicious. Otherwise, in the early stages of training, this will result in false evaluation of the legitimacy of the nodes. Moreover, in a recent review of the topic, Lyu et al. highlighted the importance of studying how to optimize the timing of deploying the defense mechanisms. They stressed the fact that when defense mechanisms are deployed to check if any adversary is attacking the FL system, the FL server requires additional computational cost. They also put forward that different types of defense mechanisms may incur different cost against different attacks. Our paper aims at addressing the challenges by proving that there must be a monitoring period during the FL process prior to the deployment of the defenses. In our analysis, we show that the time metric has an impact on the false positive and false negative rates. By studying the context of the federated learning process, the defense mechanisms must compute the time metric as per our analysis in order to provide more accurate and efficient results. We tested our suggested monitoring process and empirical results show a decrease in false positive and false negative rates and consequently, an increase in system efficiency. The key contributions of this paper are:
- A monitoring process during the early stages of training in FL for the defense approaches that are based on behavioral pattern analysis to eliminate potential malicious *workers* from the FL process.
- Evaluation of the monitoring period required in order to improve system efficiency in terms of the time cost of the iterations and computation overheads in the beginning of training for the behavioral pattern analysis approaches.
- Implementation of the monitoring process and evaluation of its appropriate time period on a behavioral pattern analysis defense applied to the distributed learning process of different benchmark systems for text and image classification (MNIST, CIFAR-10).
The rest of the paper is organised as follows: in Section [2](#sec:background){reference-type="ref" reference="sec:background"} we present a literature review on the cybersecurity of the FL process. In Section [3](#sec:methodology){reference-type="ref" reference="sec:methodology"} is described the methodology. In Sections [4](#sec:implementation){reference-type="ref" reference="sec:implementation"} and [5](#sec:Results){reference-type="ref" reference="sec:Results"}, we describe the implementation details, the simulation and provide experimental results of the monitoring process. The primary conclusions and future work are outlined in Section [6](#sec:conclusion){reference-type="ref" reference="sec:conclusion"}.
# Literature review {#sec:background}
Typically, the FL process consists of \(\mathcal{K}\) *workers* and a *chief* that trains a model *f* with \(\textbf{w}_G\) \(\in\) \(\mathbb{R}^n\) being the global parameter vector, where *n* is the dimension of the parameter space. The resulting global model is obtained by distributed training and aggregation over the \(\mathcal{K}\) *workers* holding data samples \(\mathcal{D}_i\) locally, where \(\mid\mathcal{D}_i\mid = l_i\). The total number of data samples is \(\sum\limits_{i} l_i = l\). Each *worker* keeps its data private, i.e. \(\mathcal{D}_i\) = \(\{x_1^i,..., x^i_l\}\) is not shared with the *chief* \(\mathcal{S}\) or any other *worker* \(i \neq j, \forall j \in \mathcal{K}\). The aim of the training is to generalize beyond a test dataset \(\mathcal{D}_{test}\). At each time step *t*, a random subset of *k* *workers* is selected by the *chief*. Every *worker* \(i \in \mid k \mid\) trains a local model to minimize a loss function over its own data \(\mathcal{D}_i\). The *worker* starts the local training from the global model \(\textbf{w}_G^t\) received from the *chief* and runs an algorithm such as the stochastic gradient descent for a number of \(\mathcal{E}\) epochs with a batch size of \(\mathcal{B}\). Each *worker* then obtains a local weight vector \(\textbf{w}_i\) and then computes its local model update \(\delta_i^t\) = \(\textbf{w}_i\)-\(\textbf{w}_G^t\). The local model update is sent back to the *chief*. To obtain the global model update \(\textbf{w}_G\) for the next iteration, an aggregation mechanism is used: \(\textbf{w}_G\) = \(\textbf{w}_G^t\) + \(\sum\limits_{ i \in \mid k \mid} \alpha_i \delta_i\), where \(\alpha_i\) = \(l_i/l\) and \(\sum\limits_{i} \alpha_i = 1\). Regarding the cybersecurity of the FL process, some statistical methods aim at correcting the aggregation rule implemented at the chief node. GeoMed computes the geometric median as the proposed estimator, which assumes the Byzantine ratio satisfies n \(\geqslant\) 2m + 1. Krum is another approach that assumes the Byzantine ratio satisfies n \(\geqslant\) 2m + 3. It first finds the \(n-m-2\) closest vectors in Q for each \(V_i\), and then computes a score for each vector \(V_i\). Finally, it proposes the vector \(V_i\) with the smallest score as the next update step, i.e. \(F(V_1,..., V_n) = argmin_Vi\in_Q s(V_i)\). Bulyan was originally designed for Byzantine attacks that concentrate on a single coordinate. First, it runs Krum over *Q* without replacement for \(n-2m\) time and collect the \(n-2m\) gradients to form a selection set. It then computes *F* coordinate-wise: the i-th coordinate of *F* is equal to the average of the \(n-4m\) closest i-th coordinates to the median i-th coordinate of the selection set. Bulyan has the strictest assumption as n \(\geqslant\) 4m+3 which significantly limits its practical usage. In, to defend against local model poisoning attacks, the authors generalize RONI and TRIM. Both generalized defenses remove the local models that are potentially malicious before computing the global model in each iteration of federated learning and they start as soon as training starts. One generalized defense removes the local models that have large negative impact on the error rate of the global model (inspired by RONI that removes training examples that have large negative impact on the error rate of the model), while the other defense removes the local models that result in large loss (inspired by TRIM that removes the training examples that have large negative impact on the loss). They do not consider delaying the anomaly detection by leaving a monitoring period before starting the defense. In, the authors propose a filtering mechanism, Sniper, which is conducted by the chief, to remove attackers from the global model aggregation by examining Euclidean distances between local models. Through observing the distance between models, they found models of honest users and attackers are in different cliques. Based on this, they propose a filtering defense mechanism where during every communication, the chief must run it to filter parameters updated by attackers. As a result, Sniper can recognize honest users and drop attack success rate significantly even when multiple attackers are in the federated learning system. The chief runs Sniper during every updating rather than only at the beginning of training. This can prevent attackers upload contaminated parameters at a few epochs or after a few benign sharing. Again, they do not account for a monitoring period before starting the anomaly detection. We argue that monitoring period results in a better detection performance. In, they evaluate the vulnerability of federated learning to sybil-based poisoning attacks. They propose FoolsGold, a defense that identifies poisoning sybils based on the diversity of client updates in the distributed learning process. FoolsGold changes the federated learning algorithm, relies on standard techniques such as cosine similarity, and does not require prior knowledge of the expected number of sybils. FoolsGold uses client contribution similarity and starts the defense without leaving a monitoring period. An increasing number of approaches propose to detect malicious workers by interpreting information of the workers' behaviors in order to eliminate them from the FL process. By eliminating potential malicious workers, the techniques constitute further mitigation and help accelerate the learning process. Pan et al. propose a reinforcement learning method based on a gradient aggregation agent. The method utilize the historical interactions with the workers as experience to generate reward signals for policy learning. They start the detection which they defined as a classification task with no monitoring period and report their results after a fixed number of iterations of distributed learning. Xie et al. use the loss decrease at the current iteration on the training set to rank the workers' credibility. Afterwards, the algorithm aggregates the candidates with highest scores. The score roughly indicates the level of trust in each candidate. They start their suspicion-based aggregation rule without a monitoring period and eliminate malicious workers from the process by taking the gradient estimators with the highest scores only. We want to better analyse the convergence of the FL process with regards to time when the defenses are applied in order to account for more effective and accurate malicious *worker* detection and behavior pattern analysis. To the best of our knowledge, no previous work tried to leave a monitoring period in order to evaluate if it results in a better detection performance.
# Methodology {#sec:methodology}
Defense mechanisms that are implemented at the *chief* node try to learn from the interactions with the *workers* and leverage the information received to protect the FL process in adversarial settings. The defense restores robustness by detecting malicious *workers* and eliminating them from the FL process. In the current state of information technology security practice, detection and response come only after effective monitoring. This will allow to adequately identify malicious *workers* and take appropriate response actions.
## Security settings
Figure [\[fig:cycle\]](#fig:cycle){reference-type="ref" reference="fig:cycle"} shows the security decision loop enabling adequate decision-making in terms of detection, prioritization and response to attacks. We stress the fact that to maintain an adequate security posture, the monitoring period is required especially in FL where early detection will necessary get a significant number of false positives and false negatives in the beginning of training. Monitoring has to be done for a certain period of time after which the cycle continues with the detect, decide and respond actions as per the figure. This work aims at studying the appropriate initial monitoring period of time \(\Delta\) to consider in each context of FL. We study how long do we need to monitor before starting to detect. Instead of a random value of \(\Delta\) after which evaluating the performance of the defense methods, we put forward a strategy to compute this value. If we are able to estimate the value of \(\Delta\), we will improve the efficiency of FL in adversarial settings especially in terms of time efficiency where the running time of different defenses will improve. Precisely, the time cost of the iterations in the beginning of training will be greatly improved because only monitoring will be done during a period of \(\Delta\). After this period, the defense algorithm may start to detect malicious workers and this would have brought computation overheads to lower levels. This approach is even more appreciated in a decentralized federated learning setting or in IoT networks, such as wearable devices, or smart homes where the edge device has limited resources or connectivity. In these settings, federated learning methods must be optimized to operate and train models that efficiently adapt to changes in these systems. Improving the efficiency of FL in adversarial settings especially in terms of time efficiency and computation overhead contributes to this achievement. The value of the initial monitoring period of time \(\Delta\) varies depending on many factors: the federated learning process under study (underlying case study for the training of the machine learning model), attacking patterns and defense strategy deployed. In fact, different FL processes for training deep neural networks have learning curves that follow different distributions. Moreover, different attack patterns tend to cause instability in the learning curves of the FL process. Finally, depending on the defense strategy deployed, some may perform better than others that tend to yield accuracy that distribute in a slightly wider range.
### Attack patterns
In our threat model, malicious *workers* create malicious local model updates by submitting random values sampled from a multi-dimensional Gaussian \(\mathcal{N}(\mu,\,\sigma^{2})\). We take \(\mu\) = \[0.5\...,0.5\] \(\in\) \(R^d\) and \(\sigma\) = \(2x10^6\). This attack represents an untargeted poisoning on the FL process. Moreover, we investigate different attacking patterns. Firstly, in a static attack, a fixed number of *workers* are compromised and perform continuous attacks during the entire learning process. Another attack pattern consists of a fixed number of *workers* pretending to be benign in the first fixed number of rounds and start the attack afterwards. We call this attack the pretence attack. In a randomized attack, each compromised *worker* is assigned with its role by the adversary. During the learning process, the *worker* changes its role with a certain probability.
### Defense strategy
We implement one of the defense strategies proposed in Mallah et al. for untargeted poisoning attack detection in federated learning via behavior attestation. The strategy assumes that the *chief* node has a small validation dataset. Then, the *chief* tests how the local models sent by the *worker* predict on the same validation dataset iteration after iteration in order to assess if the *worker* is reliable. The *worker*'s performance in terms of error rate compared to their own previous performance on that validation dataset will highlight if the node appears to be learning. The strategy removes the local models that have large negative impact on the error rate. The error is expected to be smaller over time when the node is not malicious. However, if the error doesn't improve nor decrease over time, then the node is considered unreliable. At iteration \(t\), for each *worker* \(j\), we compute the error, \(E_j^{t+1}\) of the current local model \(LM_j^{t+1}\) on the validation dataset. We then compare it to the error at the previous iteration, \(E_j^{t}\) as follows: \[E_j^{t+1}-E_j^{t}= E_j^{t+1}(LM_j^{t+1})-E_j^{t}(LM_j^{t})\]
## Analytical formulation
In this section, we will provide a mathematical formulation of the monitoring period. In fact, the value of the monitoring period of time \(\Delta_T\) varies depending on many factors. It may vary depending on the federated learning process under study, the attack patterns and the defense strategies. \[\Delta_T = \min(\alpha_F, \beta_A, \gamma_D)\] \[\begin{aligned}
\alpha_F &= \lfloor f(s, m, a) \rfloor\\ \beta_A &= \lfloor g(A_i) \rfloor\\ \gamma_D &= \lfloor h(D_j) \rfloor \end{aligned}\] \(\alpha_F\) is a function of the federated learning process under study *f(s, m, a)*, the underlying case study for the training of the machine learning model because different FL processes for training deep neural networks have learning curves that follow different distributions. The FL process is characterized by the size of the training set *s*, the type of model *m*, (supervised, semi-supervised or unsupervised), and the architecture *a*, (convolutional neural networks, auto-encoders, recurrent neural networks). \(\beta_A\) is a function of *\(g(A_i)\)* and corresponds to the different attack patterns (*i*= static, pretence, randomized attacks), because different patterns tend to cause instability in the learning curves of the FL process. \(\gamma_D\) is a function of *\(h(D_j)\)* and corresponds to the defense strategy deployed. Some defenses may perform better than others that tend to yield accuracy that distribute in a slightly wider range (*j*= KRUM, GeoMed, attestedFL). The monitoring period \(\Delta_T\) corresponds to the minimum number of iterations required before the performance accuracy of the model diverges from the global optima. The minimum value corresponds to the minimum of the \(\alpha_F\), \(\beta_A\), \(\gamma_D\) metrics. Training of a neural network stops when the error, i.e., the difference between the desired output and the expected output is below some threshold value or the number of iterations or epochs is above some threshold value. Thus, the value of \(\Delta_T\) ranges from 0 to the threshold value implemented. \(\Delta_T\) is equal to \(\alpha_F\) if the attack is benign and the defense is robust. This means that if an attack has a small impact on the FL process and the defense was able to neutralize it very quickly, the monitoring period required would be in line with the baseline behavior. \(\Delta_T\) will be equal to \(\beta_A\) if the attack is more powerful than the defense. Thus, \(\beta_A\) becomes the limiting factor and the monitoring period of time advised should enable the defense to start eliminating earlier the malicious nodes from the aggregation. Finally, depending on the type of defense, \(\Delta_T\) can be equal to \(\gamma_D\), thus enabling a larger monitoring period before any anomaly detection can take effect. We developed a search process that will provide an empirical validation of the analytical results. Different datasets are used to demonstrate experimentally, under different scenarios, the monitoring period in terms of delay.
# Implementation details {#sec:implementation}
We implement two use cases of federated learning with Torch, a framework for deep learning. The first use case consist in training a fully connected neural network for the hand-written digital classification task on the MNIST dataset, with 10 workers. This public dataset contains 60,000 28×28 images of 10 digits for training and 10000 for testing. The model consists of 784 inputs, 10 outputs with soft-max activation and one hidden layer with 30 ReLu units. The dimension of parameters is 25,450. The second use case is training a ResNet-18 model for the image classification on the CIFAR-10 dataset with 10 workers. This dataset contains 60,000 28×28×3 images of 10 classes of objects for training and 10000 for testing. The standard model ResNet-18 has 18 end-to-end layers and 11173962 learnable parameters in total. In all use cases, each *worker* shares a copy of the training set. We simulate the federated learning setting by sequential computation of gradients on randomly sampled mini-batches. Although any other dataset could have been chosen to validate our approach, MNIST and CIFAR-10 are used as a worldwide machine learning benchmark. Because MNIST is a labeled dataset that pairs images of hand-written numerals with the name of the respective numeral, it can be used in supervised learning to train classifiers. CIFAR-10 is an established computer-vision dataset used for object recognition. Both datasets are publicly available and are the most widely used datasets for machine learning research. Their simplicity and ease of use are what make them so widely used and deeply understood. To evaluate the performance of the FL process during the first iterations of training, we need to conduct experiments where the FL process is under normal conditions (i.e. under no attack) and report the accuracy the global model attains iteration after iteration. We will do so for the two use cases under study. Also, we need to conduct experiments where each FL process is under different attacking patterns (i.e static, pretence and randomized attacks). Finally, we conduct experiments when the defense is applied. In this case, we will also report the precision and recall of the detection mechanisms during the first iterations of the learning process and that, for different values of \(\Delta\).
## Analysis process
We present a process to investigate all possible combination of experiments to ensure a complete understanding of the FL performance during the first iterations of the learning process. The process is based on a One at A Time (OAT) analysis of the different factors affecting the accuracy of the model. In fact, the value of the initial monitoring period of time \(\Delta\) depends on three factors: the federated learning process under study, attacking patterns and defense strategy deployed. In a OAT analysis, the effect of each factor on the accuracy is investigated. To do so, factors are investigated separately and one at a time analysis of each factor have to be conducted by varying one factor at a time and keeping all other factors fixed. Thus we first conduct the experiments where the FL is under no attacks. The performance of FL in each use case will serve as a baseline for the comparison with the other experiments. We then do the experiments where there are attacks and no defenses. They serve as the upper bound for the evaluation of \(\Delta\). Afterwards we apply the defense on the three different attacks at three different values of \(\Delta\). The OAT analysis will result in a total of 26 experiments (i.e. 18 + 2 + 6 = 26). For the experiments when defenses are applied, we present the precision and recall of the methods at each iteration. Because we are experimenting with three different values of \(\Delta\), precision and recall will help estimate the appropriate value of \(\Delta\) to be considered for each use case.
# Experimental Results {#sec:Results}
The performance of malicious *workers* detection methods may be reported in terms of precision and recall as for any pattern recognition, information retrieval and classification in machine learning methods. Precision, which represents the positive predictive value, is the fraction of relevant instances among the retrieved instances. Recall, which represents the sensitivity, is the fraction of the total amount of relevant instances that were actually retrieved. In Figure [\[fig:pr\]](#fig:pr){reference-type="ref" reference="fig:pr"}, we see a detection algorithm identifying 8 malicious *workers* in an FL process containing 10 benign and 12 malicious *workers* (the relevant elements). Of the 8 identified malicious *workers*, 5 actually are malicious (True Positives, TP), while the other 3 are benign (False Positives, FP). 7 malicious *workers* were missed (False Negatives, FN), and 7 benign were correctly excluded (True Negatives, TN). The detection's precision is 5/8 (true positives / all positives) and represents how valid the results are. While its recall is 5/12 (true positives / relevant elements) and represents how complete the results are. In the context of FL, a detection technique that ensures high recall but reduces precision is better than one that increases precision but reduces recall. Greater recall increases the chances of removing benign *workers* and increases the chances of removing all malicious workers. Greater precision decreases the chances of removing benign workers but also decreases the chances of removing all malicious *workers*. In sum, it is not that bad if you remove some benign *workers* from the FL process however it is bad if you don't remove almost all malicious *workers* because of the aggregation rule. Among the strategies to evaluate the performance of the detection techniques, accuracy can be a misleading metric in the context of low ratio of malicious *workers*. On the other hand, the F-measure considers precision and recall scores. In general, it is the harmonic mean, which, for the case of two numbers, coincides with the square of the geometric mean divided by the arithmetic mean. Particularly, the \(F_1\) measure corresponds to a recall and precision that are evenly weighted and represents a special case of \(\beta\) =1 in the general \(F_\beta\) measure (for non-negative real values of \(\beta\)): \[F_\beta = (1+ \beta^2) \bullet \frac{precision\bullet recall}{\beta^2 \bullet precision +recall}\] \(F_\beta\) corresponds to the effectiveness of defense strategy with respect to a *chief* who attaches \(\beta\) times more importance to recall than precision\". In our study, we use \(\beta\)=2 where the \(F_2\) measure weights recall higher than precision.
## Empirical validation
We run the two use cases (MNIST, CIFAR-10) with the defense starting at different monitoring periods of time \(\Delta\) under different attacks. We show in Figures [\[fig:MNIST_1\]](#fig:MNIST_1){reference-type="ref" reference="fig:MNIST_1"}, [\[fig:MNIST_2\]](#fig:MNIST_2){reference-type="ref" reference="fig:MNIST_2"}, and [\[fig:MNIST_3\]](#fig:MNIST_3){reference-type="ref" reference="fig:MNIST_3"}, the accuracy of the global model during the MNIST FL process under different attacks. We notice from the figures that under a static attack on MNIST, the model's accuracy decreased to 0.75 at iteration 58. The pretence attack on MNIST resulted in an accuracy of 0.89 at iteration 58 and the randomized attack has an accuracy of 0.15 at the epoch. In this use case, we notice that the attack that has the biggest impact is the randomized attack where any *worker* can be compromised with a certain probability by the adversary at every iteration. When the defense is applied to protect against the static, pretence and randomized attack, we notice in all the figures that the accuracy of the model improves. When the defense starts at \(\Delta\) = 0, in case of a static attack on MNIST, the model's accuracy is 0.72 at iteration 58. In Figure [\[fig:MNIST_2\]](#fig:MNIST_2){reference-type="ref" reference="fig:MNIST_2"}, for the pretence attack, the accuracy improved of 0.07 at iteration 58 when \(\Delta\) = 0 compared to when no defense is applied. Similarly for the randomized attack on MNIST in Figure [\[fig:MNIST_3\]](#fig:MNIST_3){reference-type="ref" reference="fig:MNIST_3"}, the accuracy also improved of 0.07 at iteration 58 when \(\Delta\) = 0. This is due to the robustness of the federated learning that is different under fixed or variable ratios of malicious *workers*. The ratio can be fixed (static attack) or variable (pretence, randomized attacks). After *t* steps of gradient descent when the ratio is fixed, the algorithm asymptotically converge to the global optimum. This is due to the fact that when the ratio is fixed over time, the defense helps the underlying system attain a sub-optimal parameter with a certain error in a predictable number of steps. The convergence of the learning process has the optimal rate O(1/t) in Byzantium-free learning case (baseline curves on the figures). Figure [\[fig:MNIST_1\]](#fig:MNIST_1){reference-type="ref" reference="fig:MNIST_1"} shows an empirical validation of the analytical result for the MNIST dataset when the ratio of malicious *workers* is fixed. However, when the ratio of malicious *workers* varies over time as in Figures [\[fig:MNIST_2\]](#fig:MNIST_2){reference-type="ref" reference="fig:MNIST_2"} and [\[fig:MNIST_3\]](#fig:MNIST_3){reference-type="ref" reference="fig:MNIST_3"}, the defense detects different rates from iteration to iteration so the error is not predictable, thus the accuracy is not stable over the training period. On another hand, we notice that for the same attack, when the defense is applied at different periods of time \(\Delta\)=10 and \(\Delta\)=40, the accuracy of the model is different. In Figure [\[fig:MNIST_1\]](#fig:MNIST_1){reference-type="ref" reference="fig:MNIST_1"} for the static attack, at \(\Delta\)=0, the accuracy reaches 0.72 at iteration 58. However, when the defense starts at \(\Delta\)=10, the accuracy is 0.82 and at \(\Delta\)=40, the accuracy is 0.78. For this attack, we realize that between \(\Delta\)=0 and \(\Delta\)=10, the accuracy increased. This means that even if we start the defense 10 iterations later, it will not have a bad impact but it would rather improve the accuracy of the model. For the pretence attack, between \(\Delta\)=0 and \(\Delta\)=10, the accuracy increased of 0.01 and for the randomized attack, it increased of of 0.14. This shows that the best protection against attacks does not need to start right at the beginning of training but can start later and even performs better. There is an optimum amount of time where if the defense starts, the federated learning process can still be protected against attacks in an efficient way. We call this period of time, a monitoring period of duration \(\Delta\). However, we notice that for all attacks, if the defense was to be applied later, at \(\Delta\)=40, the accuracy decreases significantly compared to when it is applied at \(\Delta\)=10. For the static attack on MNIST, the accuracy decreased of 4% for the defense at \(\Delta\)=40 compared to when it was started at \(\Delta\)=10. Similarly for the pretence attack and the randomized attack on MNIST where the accuracy decreased of 7% and 14% respectively between \(\Delta\)=10 and \(\Delta\)=40 at iteration 58 in Figures [\[fig:MNIST_2\]](#fig:MNIST_2){reference-type="ref" reference="fig:MNIST_2"}, [\[fig:MNIST_3\]](#fig:MNIST_3){reference-type="ref" reference="fig:MNIST_3"}. This indicates that the defense must not start too late otherwise the training is impacted by the attacks. This result was expected because the attackers had lots of time to sabotage the system. The defense that came later was not able to increase the model's accuracy in comparison to the defense that was applied earlier and was already eliminating malicious nodes from the FL process. We notice that for this use case, the randomized attack is the most impacted by a delay before a defense is applied. Figures [\[fig:CIFAR_1\]](#fig:CIFAR_1){reference-type="ref" reference="fig:CIFAR_1"}, [\[fig:CIFAR_2\]](#fig:CIFAR_2){reference-type="ref" reference="fig:CIFAR_2"}, and [\[fig:CIFAR_3\]](#fig:CIFAR_3){reference-type="ref" reference="fig:CIFAR_3"} show the accuracy of the global model for the CIFAR-10 use case under the static, pretence and randomized attacks with the defense starting at different \(\Delta\) periods. When the defense is applied to protect against the attacks, we notice in all the figures that the accuracy of the model improves. Similar to MNIST, when the defense starts at \(\Delta\)=10 in Figure [\[fig:CIFAR_1\]](#fig:CIFAR_1){reference-type="ref" reference="fig:CIFAR_1"}, the model's accuracy increases of 0.13 at iteration 58 in comparison to when the defense starts at \(\Delta\)=0. This means that even if we start the defense 10 iterations later, it will not have a bad impact but it would rather improve the accuracy of the model. In Figure [\[fig:CIFAR_2\]](#fig:CIFAR_2){reference-type="ref" reference="fig:CIFAR_2"}, for the pretence attack on CIFAR, the accuracy improved of 0.08 at iteration 58 when the defense started at \(\Delta\) = 10 compared to when it started at \(\Delta\)=0. For the randomized attack in Figure [\[fig:CIFAR_3\]](#fig:CIFAR_3){reference-type="ref" reference="fig:CIFAR_3"}, the accuracy remained the same at iteration 58. Also, we found a similar result for CIFAR as for MNIST regarding the case when the defense is applied later, at \(\Delta\)=40. The defense must not start too late because the accuracy decreases compared to when it is applied at \(\Delta\)=10. For the static attack on CIFAR in Figure [\[fig:CIFAR_1\]](#fig:CIFAR_1){reference-type="ref" reference="fig:CIFAR_1"}, the accuracy decreased of 10% for the defense at \(\Delta\)=40 compared to when it was started at \(\Delta\)=10. For the pretence attack in Figure [\[fig:MNIST_2\]](#fig:MNIST_2){reference-type="ref" reference="fig:MNIST_2"}, the accuracy increased of 2%. For the randomized attack in Figure [\[fig:MNIST_3\]](#fig:MNIST_3){reference-type="ref" reference="fig:MNIST_3"}, the accuracy decreased of 3% between \(\Delta\)=10 and \(\Delta\)=40 at iteration 58. For this use case, the static attack is the most impacted by a delay before a defense is applied. During a randomized attack on CIFAR where the accuracy remained the same, this is still an important result in terms of improvement of system efficiency as related to time cost and computation overheads. The conclusion from the results obtained from the evaluation of the two datasets is that anomaly detection in federated learning can be delayed while the same or even better performance in terms of accuracy can be achieved. On the other hand, in terms of system performance, the monitoring period will undoubtedly decrease computation overhead and time cost. To understand the reason behind the variation in the accuracy of the global model when the defense starts at different periods of time, we must look into the precision and recall scores of the detection mechanism during the first iterations of training. This will also enable us to validate the need of a monitoring period when defending the FL process against attacks. In Figures [\[fig:MNIST_precisionrecall1\]](#fig:MNIST_precisionrecall1){reference-type="ref" reference="fig:MNIST_precisionrecall1"} and [\[fig:MNIST_precisionrecall2\]](#fig:MNIST_precisionrecall2){reference-type="ref" reference="fig:MNIST_precisionrecall2"}, we show the precision and recall in terms of percentage and the values are reported for the defense against the static attack on MNIST applied at different \(\Delta\) periods of time. We do not show the figures of the similar analysis for the pretence and randomized attacks because they show a similar behavior. When \(\Delta\) = 0, we notice from Figure [\[fig:MNIST_precisionrecall1\]](#fig:MNIST_precisionrecall1){reference-type="ref" reference="fig:MNIST_precisionrecall1"} that the value of precision is almost always higher than recall at every iteration. However, when \(\Delta\) = 10, which corresponded in Figure [\[fig:MNIST_1\]](#fig:MNIST_1){reference-type="ref" reference="fig:MNIST_1"} to the monitoring period enabling the best global model accuracy when protecting against a static attack, the values of recall are higher than precision at almost every iteration. When \(\Delta\) = 40, we notice from Figure [\[fig:MNIST_precisionrecall2\]](#fig:MNIST_precisionrecall2){reference-type="ref" reference="fig:MNIST_precisionrecall2"} that the value of precision is almost always higher than recall at every iteration. Thus, this does not represent a good indication that this must be the appropriate value of \(\Delta\). This validates the fact that in the context of FL, a detection technique that ensures high recall but reduces precision is better than one that increases precision but reduces recall. The appropriate monitoring period can be found by looking at the values of recall and precision for a subset of *workers* of the FL process. Afterwards, this period can be implemented on the the rest of the nodes of the network in order to defend against attacks in an efficient way. Similarly, Figure [\[fig:CIFAR_precisionrecall\]](#fig:CIFAR_precisionrecall){reference-type="ref" reference="fig:CIFAR_precisionrecall"} shows the precision and recall in terms of percentage for the defense against the static attack on CIFAR applied at different \(\Delta\) periods of time. The monitoring period that ensures high recall but reduces precision during training is considered the best. From the Figure, \(\Delta\)=10 can be distinguished by the high values of recall compared to the precision throughout the training interval considered. Precisely, estimating the appropriate value of \(\Delta\) must consider both precision and recall scores. We present in Figure [\[fig:MNIST_F\_beta\]](#fig:MNIST_F_beta){reference-type="ref" reference="fig:MNIST_F_beta"}, the \(F_2\) values at every iteration as a measure of the performance of the defense for different \(\Delta\) periods. The \(F_2\) measure weights recall higher than precision. Similarly, Figure [\[fig:CIFAR_F\_beta\]](#fig:CIFAR_F_beta){reference-type="ref" reference="fig:CIFAR_F_beta"} presents the \(F_2\) values for the CIFAR use case and they measure the effectiveness of detection with respect to a *chief* who attaches more importance to recall than precision. In both use cases, low values of F-score are associated to \(\Delta\) =40. This monitoring period is thus not the most appropriate for the models under study. In CIFAR, the highest values of F-score are associated to a monitoring period of \(\Delta\) =10. In fact, this monitoring period is the most efficient for the CIFAR FL process. For MNIST, F-score values of \(\Delta\) =10 are very close to the values attained when \(\Delta\) =0, so even if the defense starts at a later time, the scores are not impacted a lot by this delay in protection against attacks. Finally, there is an inverse relationship between the monitoring period and the accuracy, where it is possible to increase one at the cost of reducing the other. However, as the results imply at \(\Delta\)=10, compared to when \(\Delta\)=0 and \(\Delta\)=40, adapting the monitoring period will bring time cost and consequently computation overheads down for the defense for both the MNIST and CIFAR use cases.
# Conclusion {#sec:conclusion}
In this paper, we studied how to optimize the timing of deploying the defense mechanisms against poisoning attacks in federated learning. In fact, we proposed that a monitoring period must be accounted for particularly in defense approaches that are based on behavioral pattern analysis to eliminate potential malicious *workers* from the FL process. We showed it by conducting a comparative analysis of training traces of benign and malicious *workers* during the early stages of training. We evaluated the monitoring period required in order to improve system efficiency in terms of the time cost and computation overheads in the beginning of training for different benchmark systems for text and image classification (MNIST, CIFAR-10). Empirical results suggest in most cases, the correction lowers the false positives and false negatives and helps the distributed learning system achieve better performance in the early stage of training. To fine-tune the appropriate monitoring period throughout the distributed learning process, game theoretic research holds promise in addressing this challenge. | {'timestamp': '2021-11-30T02:25:31', 'yymm': '2109', 'arxiv_id': '2109.05385', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05385'} |
null | null |
null | null |
# Introduction
Two-dimensional ensembles of magnetic nanoparticles (MNPs) are one of the essential classes of nanostructures due to their exciting physics and various technological applications . Their importance for fundamental research stems from multiple factors such as low dimensionality and a significant contribution from increased surface atoms, resulting in intriguing magnetic properties not observed in the bulk systems . Consequently, they have tremendous prospects in diverse applications such as spintronics, ultra-high-density magnetic recording media, magnetic hyperthermia, etc. . The dipolar interaction plays an essential role in stabilizing magnetic order and determining the characteristics of such a system . The dipolar interaction is long-ranged and anisotropic; it has diverse effects on various properties of vital importance such as hysteresis, collective behaviour, blocking temperature, relaxation, etc. . Geng *et al.* systematically investigated the variation of remanence as a function of various parameters such as particle size, interparticle interaction, anisotropy in the ensemble of cobalt ferrite MNPs . They observed reduction in the remanence due to dipolar interaction. Poddar *et al.* synthesized monolayers and multilayers of magnetite MNPs and performed the ac and dc magnetic susceptibility measurements . The dipolar interaction induces spin-glass-like slowing down of magnetic relaxation in such cases. Bahiana *et al.* investigated the dipolar interaction effect on magnetic ordering in a two-dimensional array of MNPs using Metropolis Monte Carlo (MMC) technique . The dipolar interaction is found to induce antiferromagnetic coupling at low temperatures. In recent work, we investigated the hysteresis properties in ordered arrays of MNPs as a function of dipolar interaction strength, aspect ratio, external magnetic field directions using the kinetic Monte Carlo (kMC) algorithm . Interestingly, the dipolar interaction of sufficient strength induces antiferromagnetic in the square sample while ferromagnetic coupling in a highly anisotropic system. The dipolar interaction also strongly affects the ground state morphology in such a system. For example, the minimum energy state is antiferromagnetic in a square array of MNPs . In contrast, the dipolar interaction promotes the ferromagnetic arrangement of spins in a triangular lattice . In a simple cubic lattice, the minimum energy spins' configuration is dictated by antiferromagnetic coupling . While in face centered cubic arrangement of MNPs, the ground state is ferromagnetic . Moreover, the dipolar interaction also significantly affects the hysteresis loop area, one of the essential quantifiers for magnetic hyperthermia application . V. Russier *et al.* investigated the hysteresis in the two-dimensional array of MNPs . The dipolar interaction is found to have a detrimental effect on the hysteresis loop area. Xue *et al.* studied the hysteresis characteristics in an infinite two-dimensional hexagonal array of MNPs . The hysteresis loop area gets decreased because of dipolar interaction. Haase and Nowak also found a detrimental effect of dipolar interaction on the amount of heat dissipated in the hysteresis process . Takagaki *et al.* also observed a strong effect of the dipolar interaction on the hysteresis in the square arrangement of nanomagnet . Tan *et al.* also observed a decrease in hysteresis loop area and tunnel magnetoresistance in the two-dimensional MNPs arrays at temperature \(T=0\) K due to the dipolar interaction . In recent work, we also observed a strong dipolar interaction effect on the hysteresis loop area in ordered arrays of MNPs . The dipolar interaction also affects the magnetotransport properties in MNPs ensembles, an essential measure for spintronics based applications . The above discussion clearly shows that the dipolar interaction has varied effects on various systematic properties in MNPs arrays. In addition, the dipolar interaction may promote ferromagnetic or antiferromagnetic coupling between the magnetic moments depending on the interparticle separation and the geometrical arrangement of MNPs. Experimentally fabricated two-dimensional arrays of magnetic nanoparticles are generally not free from defects . One of the prominent classes of defects observed in such systems is out-of-plane disorder . In such a case, all the MNPs in the arrays do not lie in a \(xy\)-plane (say); a few of them are also scattered along the \(z\)-axis. Therefore, it is essential to understand the effect of such disorder on magnetic hysteresis and provide a theoretical basis to explain so-observed unexpected magnetic behaviour. These studies are of immense importance for efficient use of such systems in spintronics-based applications, digital data storage, and other related applications where such nanostructures are ubiquitous. Thus motivated, in the present work, we systematically investigate the effect of out-of-plane disorder strength, dipolar interaction, system size, aspect ratio and external magnetic field directions on the hysteresis mechanism in the two-dimensional MNPs ensembles using kinetic Monte Carlo simulation.
# Theoretical Framework
We consider two-dimensional arrays (\(l^{}_x\times l^{}_y\)) of spherical and monodisperse MNPs. The particles are assumed to lie in \(xy\)-plane with a fraction of them scattered along the \(z\)-axis, as shown in schematic Fig. ([\[figure1\]](#figure1){reference-type="ref" reference="figure1"}). Such positional defects are termed as out-of-plane disorder, advertently observed in experiments. Let the nanoparticle has diameter \(D\) and the lattice spacing \(l\) \[see Fig. ([\[figure1\]](#figure1){reference-type="ref" reference="figure1"})\]. The anisotropy energy \(E^{}_K\) associated with such particle is given by \[E^{}_K=K^{}_{\mathrm{eff}}V\sin^2\Phi\] Here \(K^{}_{\mathrm {eff}}\) is the anisotropy constant corresponding to uniaxial anisotropy vector \(\vec{K}= K^{}_{\mathrm {eff}} \hat{k}\), \(\hat{k}\) is the unit vector along with the anisotropy or easy axis. The nanoparticle has volume \(V=\pi D^3/6\) and \(\Phi\) is the angle between \(\hat{k}\) and magnetic moment vector. In an assembly, MNPs interact due to the dipolar interaction. The corresponding interaction energy \(E^{}_{\mathrm {dip}}\) can be calculated using the following expression \[\label{dipole} E^{}_{\mathrm {dip}}=\frac{\mu^{}_o\mu^2}{4\pi l^3}\sum_{j,\ j\neq i}\left[ \frac{\hat{\mu_{i}}\cdot\hat{\mu_{j}}-3\left(\hat{\mu_{i}}\cdot\hat{r}_{ij}\right)\left(\hat{\mu_{j}}\cdot\hat{r}_{ij}\right)}{(r^{}_{ij}/l)^3}\right].\] Where the permeability of free space is \(\mu_{o}\); \(\mu=M^{}_sV\) is the magnetic moment of a nanoparticle, \(M^{}_s\) is the saturation magnetization. The \(i^{th}\) and \(j^{th}\) magnetic moments have unit vectors \(\hat{\mu}_{i}\) and \(\hat{\mu}_{j}\), respectively, separated by a distance \(r^{}_{ij}\). \(\hat{r}^{}_{ij}\) is the unit vector corresponding to \(\vec{r}_{ij}\). The corresponding field to the dipolar interaction is expressed as \[\mu^{}_{o}\vec{H}^{}_{\mathrm {dip}}=\frac{\mu_{o}\mu}{4\pi l^3}\sum_{j,j\neq i}\frac{3(\hat{\mu}^{}_j \cdot \hat{r}_{ij})\hat{r}^{}_{ij}-\hat{\mu^{}_j} }{(r_{ij}/l)^3}. \label{dipolar1}\] We model the dipolar interaction strength by a control parameter \(h^{}_d=D^{3}/l^3\) . It correctly captures the physics of dipolar interaction variation as the latter varies as \(1/r^{3}_{ij}\), evident from Eq. ([\[dipole\]](#dipole){reference-type="ref" reference="dipole"}) and Eq. ([\[dipolar1\]](#dipolar1){reference-type="ref" reference="dipolar1"}). Therefore, the magnetically non-interacting system can be modelled with \(h^{}_d=0\) while \(h^{}_d= 1.0\) corresponds to the strongest dipolar interacting case. The alternating magnetic field is applied to the system to study the magnetic hysteresis. It is given by \[\mu^{}_{o}H=\mu^{}_oH^{}_{\mathrm {max}}\cos\omega t, \label{magnetic}\] where \(\mu^{}_{o}H_{\mathrm {max}}\) is the amplitude of the external magnetic field; \(\omega=2\pi\nu\) is the angular frequency. \(\nu\) is the linear frequency, and \(t\) is time. We can therefore write the expression for the total energy \(E\) as \[E=K^{}_{\mathrm {eff}}V\sum_{i}\sin^2 \Phi^{}_i+\frac{\mu^{}_o\mu^2}{4\pi l^3}\sum_{j,\ j\neq i}\left[ \frac{\hat{\mu_{i}}\cdot\hat{\mu_{j}}-{3\left(\hat{\mu_{i}}\cdot\hat{r}_{ij}\right)\left(\hat{\mu_{j}}\cdot\hat{r}_{ij}\right)}}{(r_{ij}/l)^3}\right]-\mu^{}_o\mu\sum_{i}\hat{\mu}^{}_i\cdot\vec{H}\] Here \(\Phi^{}_i\) is the angle between the anisotropy axis and the \(i^{th}\) magnetic moment of the system. We use state of the art kinetic Monte Carlo simulation to investigate the magnetic hysteresis in the two-dimensional assembly of MNPs in the presence of the out-of-plane disorder. To quantify the out-of-plane-disorder, we define a disorder strength parameter \(\Delta(\%)=n^{}_z/N\), where \(N\) is the total number of nanoparticles in the system and \(n^{}_z\): the number of particles scattered along the \(z\)-axis (normal to the plane of the system). In the present work, we systematically analyzed the hysteresis response as a function of \(\Delta (\%)\), dipolar interaction strength \(h^{}_d\), aspect ratio \(A^{}_r=l_y/l_x\), and the external magnetic field's direction. The kMC algorithm is a more suitable choice than the MMC technique to investigate the time-dependent properties such as dynamic hysteresis, relaxation characteristics, etc. It is because the dynamics are well captured in the kMC methods . Tan *et al.* implemented the same kMC procedures to investigate the spatial distribution of heat dissipation in the spherical assembly of MNPs . We have also used it to study various properties in arrays of dipolar interacting nanoparticles . The area under the hysteresis curve equals the amount of heat disspated \(E^{}_{H}\) during one complete cycle of external field variation, which can be evaluated using the following formula \[E^{}_{H}=\oint M(H)dH, \label{local_heat}\] The above intergral is evaluated over the entire period of the external magnetic field change. \(M^{}(H)\) is the magnetization of system at the applied magnetic field \(H\).
# Simulations Results
The nanoparticles are assembled in the two-dimensional (\(l^{}_x\times l^{}_y\)) lattice \[\(xy\)-plane\] with fraction of them scattered along \(z\)-direction. We have considered following values of simulations parameters: \(D=8\) nm, \(K^{}_{\mathrm {eff}}=13\times10^{3}\) \(\mathrm {Jm^{-3}}\), \(M^{}_s=4.77\times10^{5}\) \(\mathrm {Am^{-1}}\), \(N=400\), \(\mu^{}_oH^{}_{\mathrm {max}}=0.10\) T, \(\nu=10^5\) Hz, and \(T=300\) K. Five typical values of system sizes viz. \(l^{}_x\times l^{}_y=20\times20\), \(10\times40\), \(4\times100\), \(2\times200\), and \(1\times400\) are considered. These correspond to \(A^{}_r(=l^{}_y/l^{}_x)=1.0\), 4.0, 25.0, 100.0, 400.0, respectively. We have varied \(\Delta (\%)\) from 5 to 50 to extensively investigate the effect of out-of-plane positional disorder on the hysteresis characteristics. The dipolar interaction strength \(h^{}_d\) is varied from 0 to 1.0. We have also assumed anisotropy axes to be randomly oriented in three-dimensional space to mimic the real system. The oscillating magnetic field is applied along \(x\) and \(y\)-direction. We first investigate the out-of-plane disorder effect on the hysteresis in square arrays of MNPs as a function of dipolar interaction strength. In Fig. ([\[figure2\]](#figure2){reference-type="ref" reference="figure2"}), we plot the magnetic hysteresis curve as a function of \(\Delta (\%)\), \(h^{}_d\), and external magnetic field's direction with \(A^{}_r=1.0\) (\(l^{}_x\times l^{}_y=20\times20\)). The alternating magnetic field is applied along \(x\) and \(y\)-direction. The hysteresis curve without the disorder (\(\Delta(\%)=0\)) is also plotted for comparison. In all the hysteresis curves shown in the present work, the external magnetic field (\(x\)-axis) and the magnetization (\(y\)-axis) is rescaled by the single-particle anisotropy field \(H^{}_K=2K^{}_{\mathrm {eff}}/M^{}_s\) and \(M^{}_s\), respectively. The superparamagnetic character is dominant with weak and negligible dipolar interaction strength \(h^{}_d\leq0.2\), irrespective of \(\Delta\) and applied field's directions. Consequently, a characteristic hysteresis curve for superparamagnetic particles (S-shaped hysteresis curve with the negligible value of coercive field and remanence) is observed in such a case. Nowak *et al.* also reported similar results using MMC simulations . Interestingly, the double-loop hysteresis emerges with appreciable interaction strength \(h^{}_d=0.4\) and \(0.6\) in the absence of disorder (\(\Delta(\%)=0\)), reminiscent of antiferromagnetic coupling dominance. It can be explained by carefully analyzing the nature of the dipolar and anisotropy field as follows. The dipolar interaction forces the magnetic moments to align in the plane of the sample by inducing a biasing field in the planar arrangement of MNPs. In contrast, the anisotropy field instigates the magnetic moments to orient in three-dimensional space as the anisotropy axes are considered to be randomly oriented in the present article. As a result, there is a linear variation of magnetization with the external field, provided it is weaker than the coercive field, i.e. \(\mu^{}_o H/H^{}_K< 0.5\) . The dipolar field dictates the hysteresis behaviour for \(\mu^{}_o H/H^{}_K>0.5\), resulting in a spontaneous orientation transition of magnetization. These results are in perfect agreement with the work of Yang *et al.* . It is also clearly evident that there is a strong effect of disorder on hysteresis response. Remarkably, the out-of-plane disorder puts the system from an antiferromagnetic to a ferromagnetic magnetic state with the field applied along the \(y\)-direction. Consequently, the hysteresis curve starts to open up near the small external magnetic field (coercive field), indicating enhancement of ferromagnetic coupling in the system. Therefore, the area under the hysteresis curve increases with an increase in disorder strength \(\Delta\). Next, we analyze the magnetic hysteresis as a function of dipolar interaction and disorder strength in rectangular arrays of MNPs. Fig. ([\[figure3\]](#figure3){reference-type="ref" reference="figure3"}) shows the hysteresis curve as a function of \(\Delta\), \(h^{}_d\) and applied field directions with \(A^{}_r=4.0\) (\(l^{}_x\times l^{}_y=10\times40\)). All other parameters are the same as that of Fig. ([\[figure2\]](#figure2){reference-type="ref" reference="figure2"}). In the presence of small \(h^{}_d=0.2\), the superparamagnetic character is dominant in this case also. Notably, the non-hysteresis is observed with the field applied along the shorter axis of the sample (\(x\)-axis) for appreciable dipolar interaction strength (\(h^{}_d>0.2\)) and significant \(\Delta\). The dipolar interaction promotes ferromagnetic coupling with \(\Delta\) and \(h^{}_d\) along the long axis of the system (\(y\)-axis) because of its anisotropic nature. Consequently, the magnetization ceases to follow the external field when applied along the shorter axis (\(x\)-axis), resulting in a minimal hysteresis loop area. Interesting physics unfolds when the field is applied along the long axis of the system. The double-loop hysteresis emerges in an ordered ensemble of MNPs (negligibly small \(\Delta\)) with appreciable \(h^{}_d\), characteristics of antiferromagnetic coupling. This observation is in perfect agreement with the work of Chen *et al.* . Remarkably, the system's rapid transition from antiferromagnetic to the ferromagnetic regime with \(\Delta\), provided \(h^{}_d\) is significant. Because the dipolar interaction increases the ferromagnetic coupling with an enhancement in the disorder strength, resulting in an enhanced hysteresis loop area. These observations are also in perfect agreement with the work of Cheng *et al.* . These results clearly suggest that we can tune the nature of dipolar interaction from ferromagnetic to antiferromagnetic or vice versa by manoeuvering the disorder strength \(\Delta\) and \(h^{}_d\). These observations could be beneficial in manipulating the hysteresis loop area and other related properties of interest, which could be helpful in various applications such as magnetic hyperthermia and spintronics. After that, we investigate the hysteresis mechanism in systems with large aspect ratios. In Fig. ([\[figure4\]](#figure4){reference-type="ref" reference="figure4"}) and Fig. ([\[figure5\]](#figure5){reference-type="ref" reference="figure5"}), we plot the hysteresis curve with \(A^{}_r=25.0\) and \(A^{}_r=100.0\), respectively. All the parameters are the same as that of Fig. ([\[figure3\]](#figure3){reference-type="ref" reference="figure3"}). There is a negligible value of coercive field and remanence for weakly interacting nanoparticles, indicating superparamagnetic behaviour, irrespective of Ar and external field directions. The magnetization ceases to follow the external field (\(\mu^{}_o\vec{H}=H^{}_o\hat{x}\)) for appreciable \(h^{}_d\), resulting in non-hysteretic behaviour. Notably, the dipolar interaction induces shape anisotropy or ferromagnetic coupling along the long axis with significant \(A^{}_r\). Consequently, remanence and coercive field are significant even in the perfectly ordered system (\(\Delta(\%)=0\)) with the external magnetic field along the long axis of the sample, i.e. \(y\)-axis. Moreover, the hysteresis loop area is also significant and increases with \(h^{}_d\). There is a weak dependence of hysteresis on the out-of-plane disorder as anticipated. The ferromagnetic coupling strength has already at its maximum even with \(\Delta(\%)=0\) as the system is highly anisotropic. These observations clearly indicate that dipolar interaction of sufficient strength puts the underlying system from superparamagnetic to the ferromagnetic regime even at \(T=300\) K and \(\Delta(\%)\)=0 (ordered arrays). Therefore, there is a huge hysteresis loop area with the external field along the \(y\)-direction. In Fig. ([\[figure6\]](#figure6){reference-type="ref" reference="figure6"}), we study the hysteresis response with a highly anisotropic system \(A^{}_r=400\), which corresponds to a one-dimensional chain of MNPs. MNPs stays in the superparamagnetic regime for small dipolar interaction strength even with huge \(A^{}_r\), resulting in the minimal hysteresis loop area. We observed non-hysteresis with the field applied along the shorter axis of the sample (\(x\)-axis) with significant \(h^{}_d\). These observations are independent of the disorder and dipolar interaction strength. Remarkably, the hysteresis loop area enhances as \(h^{}_d\) increases, provided the field is applied along the long axis of the system (\(y\)-axis). The dipolar interaction promotes ferromagnetic coupling in the highly anisotropic system, such as the linear and cylindrical arrangement of MNPs. Varón *et al.* also found the similar behaviour of magnetic moments in low-dimensional system . Consequently, there is a natural tendency of magnetic moments to get aligned along the long axis of the sample even in the absence of an external field, resulting in a huge remanence(\(\sim 1.0\)) and hysteresis loop area. The increase of hysteresis loop area with \(h^{}_d\) is also in the perfect agreement with our recent work . It is also essential to investigate the hysteresis mechanism quantitatively by analyzing the coercive field \(\mu^{}_oH^{}_c\), remanence \(M^{}_r\) and amount of heat dissipation \(E^{}_H\) (hysteresis loop area) as a function of these parameters to have a better understanding of the hysteresis characteristics. Fig. ([\[figure7\]](#figure7){reference-type="ref" reference="figure7"}) and Fig. ([\[figure8\]](#figure8){reference-type="ref" reference="figure8"}) show the variation of \(\mu^{}_oH^{}_c\) and \(M^{}_r\) as a function of \(\Delta(\%)\) and \(h^{}_d\), respectively. We have considered four typical values of \(A^{}_r=1.0\), 4.0, 100.0, and 400.0, and the external magnetic field is applied along the \(y\)-axis as exciting physics unfolds in such a case. \(\mu^{}_oH^{}_c\) and \(M^{}_r\) are negligibly small with weakly dipolar interacting MNPs (\(h^{}_d\leq0.2\)), irrespective of \(\Delta\) and \(A^{}_r\), indicating the superparamagnetic character. In the absence of disorder (\(\Delta\approx0\)), \(\mu^{}_oH^{}_c\) and \(M^{}_r\) have small values even with large \(h^{}_d\) and more minor \(A^{}_r\). It is because the dipolar interaction promotes antiferromagnetic coupling among the magnetic moments in the square-like arrangement of MNPs. On the other hand, there is an enhancement in \(M^{}_r\) and \(\mu^{}_oH^{}_c\) as dipolar interaction strength \(h^{}_d\) increases for enormous \(A^{}_r\) and fixed disorder strength. We can explain it by the fact the ferromagnetic interaction is promoted due to dipolar interaction in the highly anisotropic system (huge \(A^{}_r\)). Notably, the remanence \(M^{}_r\) is huge \(\sim 1.0\) even with moderate interaction strength (\(h^{}_d\approx 0.6\)) in a system with a large aspect ratio. The dipolar interaction induces shape anisotropy along the long axis of the system (\(y\)-axis in the present case) with huge \(A^{}_r\), instigating magnetic moments to get aligned along \(y\)-direction even in the absence of an external magnetic field. The increase of \(M^{}_r\) and \(\mu^{}_oH^{}_c\) with \(h^{}_d\) in the absence of \(\Delta\) is perect agreement with the work of Figueiredo *et al.* . We could not compare our results in the presence of defects with them as they have focused only on the pefectly ordered MNPs ensembles. Finally, we study the variation of the heat dissipation \(E^{}_H\) due to hysteresis as a function of disorder strength \(\Delta\) and dipolar interaction strength \(h^{}_d\) in Fig. ([\[figure9\]](#figure9){reference-type="ref" reference="figure9"}). We have considered four representative values of aspect ratio \(A^{}_r=1.0\), 4.0, 100.0, and 400.0. \(E^{}_H\) has minimal value for weakly dipolar interacting MNPs, irrespective of \(\Delta\) and \(A^{}_r\), reminiscent of superparamagnetic behaviour. In a perfectly ordered state (\(\Delta=0\)), the antiferromagnetic coupling is dominant with significant \(h^{}_d\) and relatively smaller \(A^{}_r\). Consequently, the amount of heat dissipated \(E^{}_H\) is tiny in these cases \[see Fig. [\[figure9\]](#figure9){reference-type="ref" reference="figure9"}(a)-(b)\]. Remarkably, \(E^{}_H\) increases with \(\Delta\) in the presence of appreciable \(h^{}_d\) and smaller \(A^{}_r\). It is because the dipolar interaction of sufficient strength puts the system from antiferromagnetic to ferromagnetic state with an increase in out-of-plane disorder strength \(\Delta\). On the other hand, \(E^{}_H\) increases with \(h^{}_d\) for a given \(\Delta\) and huge \(A^{}_r\) because of an enhancement in ferromagnetic coupling. Moreover, \(E^{}_H\) depends weakly on out-of-plane disorder strength \(\Delta\) in a such case as ferromagnetic interaction is already attained its maximum.
# Summary and conclusion
Two-dimensional arrays of MNPs are of immense importance due to their usages in various applications such as digital data storage, magnetic hyperthermia and spintronics, etc. Generally, the experimentally fabricated assays are not perfectly ordered . In such cases, it is quite possible that few of the nanoparticles lie off-axis, normal to the sample plane . Such positional defects are termed out-of-plane disorders. It is also a well-known fact that dipolar interaction is greatly affected by nanoparticles's interparticle separation and geometrical arrangement. Therefore, it is essential to investigate the magnetic properties as a function of out-of-plane disorder strength \(\Delta\), dipolar interaction strength \(h^{}_d\) and other system parameters such as aspect ratio \(A^{}_r\) of the system. In the present work, we analyzed the hysteresis characteristics in the two-dimensional arrays (\(l^{}_x\times l^{}_y\)) as a function of \(\Delta\), \(h^{}_d\), \(A^{}_r=l^{}_y/l^{}_x\), and external magnetic field directions using kinetic Monte Carlo simulations. We have implemented the kinetic Monte Carlo simulation algorithm for these studies, which is superior to the Metropolis Monte Carlo technique for dynamical properties analysis. In the presence of weak dipolar interaction (\(h^{}_d\leq0.2\)), the hysteresis is dictated by superparamagnetic behaviour, resulting in the minimal hysteresis loop area. Our observation is in the perfect agreement with the work of Sousa *et al.* . Remarkably, the dipolar interaction of sufficient strength promotes antiferromagnetic coupling in a square-like well ordered (\(\Delta\approx0\)) MNPs arrays. Consequently, the double hysteresis loop emerges, characteristics of antiferromagnetic coupling dominance. Exciting physics appears with the introduction of the out-of-plane disorder in the underlying system. Yang *et al.* also found the antiferromagnetic interaction dominance in nanoparticle arrays . Interestingly, the dipolar interaction of enough strength drives the system from antiferromagnetic to ferromagnetic magnetic state in the presence of large \(\Delta\) even with smaller \(A^{}_r\) (square assembled of MNPs). On the other hand, an additional anisotropy (shape anisotropy) is developed due to the dipolar interaction in the highly anisotropic system along the long axis of arrays (\(y\)-axis in the present case). As a consequence, the magnetic moments tend to align ferromagnetically, resulting in an enhancement in the hysteresis loop area with the field applied along the \(y\)-direction. The increase of area under hysteresis curve with interaction strength is in perfect qualitative agreement with the work of Kechrakos *et al.* . In contrast, non-hysteresis is observed when the external field is applied shorter axis of the system, i.e. \(x\)-axis. These results suggest that we can tune the nature of dipolar interaction from antiferromagnetic to ferromagnetic by just inducing positional disorder even in the square assembly of MNPs. Therefore, these observations provide a concrete theoretical basis to manipulate the nature of dipolar interaction in a more controlled manner. We are also able to explain the observed hysteresis behaviour of such experimentally obtained MNPs assays. The quantitative analysis of coercive field \(\mu^{}_oH^{}_c\), remanence \(M^{}_r\) and the heat dissipation \(E^{}_H\) variation with these parameters could help assess the hysteresis properties. In the presence of weak or negligible dipolar interaction, \(\mu^{}_oH^{}_c\) and \(M^{}_r\) are minimal irrespective of \(A^{}_r\) and applied magnetic field directions, indicating superparamagnetic behaviour. Remarkably, \(M^{}_r\) and \(\mu^{}_oH^{}_c\) are extremely small with large \(h^{}_d\) and smaller \(A^{}_r\), quantifying the antiferromagnetic coupling dominance. Interestingly, there is an enhancement in these values as \(\Delta\) is increased, provided \(h^{}_d\) is large enough. The nature of dipolar interaction changes from antiferromagnetic to ferromagnetic in the presence of significant out-of-plane disorder strength \(\Delta\) with relatively more minor \(A^{}_r\). In the highly anisotropic system (huge \(A^{}_r\)), \(\mu^{}_oH^{}_c\) and \(M^{}_r\) have enormous values even in the presence of moderate \(h^{}_d\) with the field applied along the long axis of the system. In contrast, they are minimal when the field is along the shorter axis of the sample, indicating non-hysteretic behaviour. Notably, \(E^{}_H\) varies with these parameters exactly same manner as that of the coercive field, implying an intimate relationship between them. In conclusion, we analyse the hysteresis properties in the two-dimensional arrays of MNPs with the out-of-plane disorder to provide a theoretical basis for unusual characteristics in such systems. The role of dipolar interaction is strongly affected by the disorder strength, aspect ratio of the system. In particular, the dipolar interaction of enough strength drives the magnetic order from antiferromagnetic to ferromagnetic in the square-like arrangement of MNPs. While in the highly anisotropic system, even the moderate dipolar interaction is able to induce ferromagnetic coupling among the magnetic moments. The results obtained in the present work should help the experimentalist to explain the physics of the observed hysteresis response for the nanofabricated MNPs assays. These are also useful in manipulating the nature of dipolar interaction in a more controlled manner analytically. Therefore, we hope that the present work could pave the way for joint efforts in experimental, theoretical, and computational studies for these excellent and versatile systems. | {'timestamp': '2021-09-14T02:14:58', 'yymm': '2109', 'arxiv_id': '2109.05323', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05323'} |
# Introduction
The *3D-index* is a fascinating invariant of ideal triangulations of oriented 3-manifolds with toroidal boundary which associates a \(q\)-series with integer coefficients to each first homology class of the boundary. It was originally obtained by Dimofte, Gaiotto and Gukov in a mathematical physics context from a certain superconformal field theory. Garoufalidis provided an axiomatic formulation of the 3D-index and studied convergence of this invariant, which is only defined for some ideal triangulations. Although the physics predicts that the 3D-index should be a topological invariant, there is no *a priori* mathematical reason that this should be so. While the 3D-index is invariant under Pachner 2-3 moves on ideal triangulations (whenever it is defined on both sides of the move), it is not known whether all topological ideal triangulations on which the 3D-index is defined can be connected by such moves, so topological invariance does not follow. Nonetheless, Garoufalidis, Rubinstein, Segerman and the first named author showed in that the 3D-index does indeed give a topological invariant in the special case of cusped hyperbolic 3-manifolds. This paper concerns a certain extension of the 3D-index which was introduced by Garoufalidis and Kashaev and which we will call the *meromorphic 3D-index*. The extension associates to *every* ideal triangulation of an oriented 3-manifold with toroidal boundary a meromorphic function which has 2 variables for every boundary component. The meromorphic 3D-index is invariant under *all* 2-3 Pachner moves, and is hence a topological invariant. The meromorphic 3D-index is an extension of the \(q\)-series 3D-index in the following sense. In the case that the ideal triangulation has a strict angle structure, there is a Laurent expansion of the meromorphic 3D-index whose coefficients are the \(q\)-series appearing in the original 3D-index. Therefore, we can think of the meromorphic 3D-index as a kind of generating function for these \(q\)-series. In particular, the \(q\)-series 3D-index is a true topological invariant when restricted to ideal triangulations admitting strict angle structures. The meromorphic 3D-index introduced in depends on a complex parameter \(q\) satisfying \(0<|q|<1\). The main goal of this paper is to study the asymptotics of this invariant when \(q\to 1\). More precisely, we express the meromorphic 3D-index in terms of a parameter \(\hbar\in (-\infty,0)\) which determines a \(q\in(0,1)\) by \(q=e^{\hbar}\). The meromorphic 3D-index evaluated on the ideal triangulation \(\mathcal{T}\) with a single torus boundary component is then denoted \(\I{\mathcal{T}}{A_\mu}{A_\lambda}(\hbar)\) and is a meromorphic function of two complex variables \(A_\mu,A_\lambda\). These variables can be interpreted as the *angle-holonomies,* along the curves \(\mu\) and \(\lambda\), of a complexified angle structure on \(\mathcal{T}\), in a sense that will be made more precise later. Of particular interest is the behaviour of this invariant at the point \(A_\mu=A_\lambda=0\), corresponding to trivial peripheral holonomies, since the value at this point does not depend on the choice of the curves \(\mu\), \(\lambda\). Moreover, in the case the manifold is hyperbolic, this point corresponds to an angle structure derived from the complete hyperbolic structure. In this paper, we mostly focus on the case of a one-cusped hyperbolic manifold \(M\) and study the asymptotic behaviour of the meromorphic 3D-index for \(\hbar=-1/\kappa\), which corresponds to \(q=\exp(-1/\kappa)\): \[\label{limit-we-study} \I{M}{0}{0}(-1/\kappa),\quad \kappa>0,\quad \kappa\rightarrow\infty.\] In Section [\[the-conjecture\]](#the-conjecture){reference-type="ref" reference="the-conjecture"}, we provide a conjectural asymptotic approximation for the above quantity, in the form of a sum over conjugacy classes of irreducible, boundary-parabolic representations \(\rho:\pi_1(M)\rightarrowPSL(2,\CC)\) with a certain fixed obstruction class to lifting the representation to \(SL(2,\CC)\). Moreover, we will use heuristic saddle-point methods to develop explicit formulae for these contributions. Our asymptotic approximations show strong agreement with numerical experiments, which we describe in Section [\[sec:numerical\]](#sec:numerical){reference-type="ref" reference="sec:numerical"}. Although the overall picture is conjectural and based on a heuristic analysis, we will present many rigorous theorems about the various expressions that arise as coefficients in our asymptotic approximations. In particular, we expect some of these coefficients to be given by classical quantities, such as the volume of a representation and the Reidemeister torsion associated to its adjoint. Moreover, we use the total predicted contribution of boundary-parabolic *real* representations \(\rho:\pi_1(M)\rightarrowPSL(2,\mathbb{R})\) to construct a new topological invariant, which we call the *beta invariant*. The beta invariant is defined in terms of state-integrals on ideal triangulations built from the Euler beta function and shown, in Section [\[sec:Mellin-Barnes\]](#sec:Mellin-Barnes){reference-type="ref" reference="sec:Mellin-Barnes"}, to be fully invariant under Pachner 2-3 moves. Along the way, we study the cohomological obstruction classes to lifting boundary-parabolic \(PSL(2,\CC)\)-representations of \(\pi_1(M)\) to boundary-unipotent \(SL(2,\CC)\)-representations and develop a theory to explain their relationship to circle-values angle structures on ideal triangulations.
## A sketch of the meromorphic 3D-index {#sketch-of-the-index}
Here we'll give a brief sketch of the construction of meromorphic 3D-index in the special case of a 3-manifold with a *single* torus boundary component. Section [\[section-index\]](#section-index){reference-type="ref" reference="section-index"} below gives a detailed review. To be precise, we need to consider two cases: the case when the triangulation admits a strict angle structure, and the case when it does not. (A strict angle structure is an assignment of strictly positive real numbers to normal quadrilateral types which sums to \(\pi\) in each tetrahedron and to \(2\pi\) around each edge.) In the case the triangulation does admit a strict angle structure, we will write down a *state-integral* expression which depends on the triangulation and a strict angle structure on it. The strict angle structure ensures convergence of the integral, but the value of the integral depends on the strict angle structure only through the peripheral angle-holonomies \(A_\mu\) and \(A_\lambda\). Thus, the state-integral determines a germ of a meromorphic function defined initially at suitable points \((A_\mu,A_\lambda)\in\mathbb{R}^2\). The full function \(\I{\mathcal{T}}{A_\mu}{A_\lambda}(\hbar)\) is then obtained as the analytic continuation of this germ into the complex domain. The other case, when the triangulation does not admit a strict angle structure, is detailed in Section [\[section-index\]](#section-index){reference-type="ref" reference="section-index"} and involves a more intricate analytic continuation argument, first presented in. In order to describe the state-integral depending on a given oriented ideal triangulation \(\mathcal{T}\) with \(N\geq 2\) tetrahedra, we shall denote by \(Q(\mathcal{T})\) the set of normal quadrilateral types in \(\mathcal{T}\) and by \(\mathcal{E}(\mathcal{T})\) the set of edges of \(\mathcal{T}\). Suppose \(\alpha: Q(\mathcal{T})\to(0,\pi)\) is a strict angle structure on \(\mathcal{T}\). For every edge \(E\in\mathcal{E}(\mathcal{T})\), there will be one integration variable \(y_E\) taking values in the unit circle \(S^1\subset\mathbb{C}\). The integrand will be a product of *tetrahedral weights*, one for each tetrahedron. The weight of each tetrahedron is a function of the integration variables appearing on the edges of that tetrahedron. Moreover, each tetrahedral weight consists of four factors--one for each normal quad type in the tetrahedron, and one additional normalisation factor. So the general shape of the state integral is \[\label{eq:general-state-integral} c_q^N \int_{(S^1)^{\mathcal{E}(\mathcal{T})}} \prod_{\square\in Q(\mathcal{T})} F(\square)\, \bigwedge_{E\in \mathcal{E}(\mathcal{T})} \mathrm{d}y_E.\] | {'timestamp': '2021-11-16T02:40:46', 'yymm': '2109', 'arxiv_id': '2109.05355', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05355'} |
null | null |
# Introduction
All three major satellite missions designed to study the cosmic microwave background (CMB)---COBE (Hinshaw et al. 1996), WMAP (Bennett et al. 2003), and *Planck* (Planck Collaboration VII 2020)---have uncovered several unexpected features in its temperature fluctuations. These include missing correlations at large angles and an unexpectedly low power in multipoles \(2\lesssim \ell\lesssim 30\), especially \(\ell=2\) and \(3\). In spite of these apparent deficiencies, however, the agreement between the angular power spectrum predicted by \(\Lambda\)CDM and the observations is quite remarkable for \(\ell>30\), forming the basis for the optimization of many cosmological parameters. The large-scale anomalies stand in sharp contrast to the general level of success interpreting the CMB anisotropies, sustaining a simmering debate concerning their origin, or possible misidentification due to unknown systematic effects (see Bennett et al. 2011, Copi et al. 2010 for reviews). If these two unexpected features are real, it is not even clear if they are due to the same physical process, in spite of the fact that they both characterize variations in temperature over very large angular scales. A thorough analysis of these anomalies was carried out with the final release of the *Planck* data by the Planck Collaboration VII (2020). They studied the statistical isotropy and Gaussianity of the CMB using *both* the Planck 2018 temperature and polarization data. Previous work had focused solely on the temperature fluctuations, due to only a limited ability to also probe the CMB polarization. The large-angular-scale polarization measurements permit a largely independent examination of the peculiar features seen in the temperature which can, in principle, reduce or eliminate any subjectivity or ambiguity in interpreting their statistical significance. The possibility that the lack of large-angle correlations in the temperature may be largely due to a statistical 'fluke,' as unlikely as that may appear to be, was raised by the fact that the Planck Collaboration VII study found only weaker evidence of an analogous lack of large-scale angular correlations in the polarization data. Nevertheless, in spite of the fact that no unambiguous detections of cosmological non-Gaussianity, or of anomalies analogous to those seen in temperature, could be claimed, efforts at reducing the systematic effects that contaminated the earlier polarization maps on large angular scales could not completelyexclude the presence of residual systematics that limit some tests of non-Gaussianity and isotropy. A comparison of both the temperature and polarization analyses thus tends to mitigate our ability to conclude one way or the other whether the large-angle anomalies are real. From a theoretical standpoint, the polarization fluctuations also originated in the primordial gravitational potential, but are largely sourced by different modes than the temperature anisotropies. Thus, a measurement of large-scale anomalies in *both* the temperature and polarization has the potential of providing a greater statistical significance. It is important to see whether any anomalies seen in the polarization maps are related to known features in the temperature. On the other hand, if no such anomalies are seen in the polarization it might be the case that the temperature anomalies, if not pure flukes, could be due to secondary influences, such as the integrated Sachs-Wolfe effect (Planck Collaboration XXI 2016), or more exotic scenarios in which they may be due to physical processes that do not correlate with the polarization. The situation with regard to the large-angle anomalies is thus far from clear, and it is essential to continue broadening the study of possible causes of such features in either the temperature, polarization or both. In earlier work (Melia & López-Corredoira 2018), we showed that a likely theoretical explanation for the missing large-angle correlations in temperature is the presence of a cutoff, \(k_{\rm min}\), in the primordial power spectrum, \({\mathcal{P}}(k)\). Indeed, our analysis of the *Planck* data revealed that a zero cutoff is ruled out by these measurements at a confidence level \(\gtrsim 8\sigma\). Instead, we found that the observed angular correlation function, \(C(\theta)\), could be reproduced satisfactorily at all angles with a cutoff \(k_{\rm min}=(4.34\pm0.50)/r_{\rm dec}\), where \(r_{\rm dec}\) is the comoving distance between us and the last scattering surface at decoupling (\(z_{\rm dec}=1080\)). A truncation such as this is not necessarily consistent with the aims of slow-roll inflation to simultaneously fix the horizon problem and seed a quantum fluctuation spectrum consistent with large-scale structure (Liu & Melia 2020, Melia 2020), because \(k_{\rm min}\) represents the time at which inflation could have started, severely constraining the number of e-folds available for the Universe to expand at an accelerated rate. It is therefore imperative to pursue this line of inquiry further and see if additional evidence may be found in favor of a non-zero \(k_{\rm min}\). The goal of this paper is to complete that analysis in two significant ways: (1) to examine whether the same cutoff may be responsible for both large-angle anomalies, and (2) to confirm that the introduction of \(k_{\rm min}\) in \({\mathcal{P}}(k)\) does not adversely impact the remarkable agreement between theory and observations at \(\ell>30\). In § 2 we introduce the basic theoretical background underlying this work, and then proceed to study the effects of a truncated \({\mathcal{P}}(k)\) on the angular power spectrum in § 3. We discuss the results in § 4, and present our conclusion in § 5.
# Theoretical background
The inflationary paradigm posits that an early accelerated expansion persisted long enough to create a homogeneous and isotropic Universe within our current Hubble volume. As such, the CMB temperature, \(T\), measured in some direction \(\hat{e}\), is expected to be a Gaussian random field on the sky, expandable as a series of spherical harmonics, \(Y_{\ell m}(\hat{e})\), with independent Gaussian random coefficients \(a_{\ell m}\) of zero mean: \[\langle a_{\ell m}\rangle=0\;,\] for all \(\ell>0\) and \(m=-\ell,-\ell+1...,+\ell\). If the Universe is indeed statistically isotropic, this spherical harmonic decomposition, and the two-point angular power spectrum derived from it, contain all of the physical information needed to interpret the CMB anisotropies. The less used (but equally important) angular correlation function \(C(\theta)\) relates the temperature measured in two independent directions, \(\hat{e}_1\) and \(\hat{e}_2\), and depends solely on the dot product \(\cos\theta\equiv\hat{e}_1\cdot\hat{e}_2\). One may thus expand it in terms of Legendre polynomials: \[\langle T(\hat{e}_1)T(\hat{e}_2)\rangle\equiv C(\theta)={1\over 4\pi}\sum_\ell (2\ell+1)\,C_\ell P_\ell(\cos\theta)\;,\label{eq:angcorr}\] in which the variance \(C_\ell\) is known as the angular power of multipole \(\ell\). Statistical independence implies that the expectation of a product of \(a_{\ell m}\)'s with different \(\ell\)'s and \(m\)'s vanishes, while isotropy requires the constant of proportionality for this product to be solely a function of \(\ell\), not \(m\): \[\langle a_{\ell m}^*\,a_{{\ell}^{\prime} {m}^{\prime}}\rangle= \delta_{\ell {\ell}^{\prime}}\,\delta_{m m^\prime}\,C_\ell\;.\] As one can see from figure [\[f1\]](#f1){reference-type="ref" reference="f1"} (discussed below), the standard model has been strikingly successful in accounting for the angular power spectrum (i.e., \(C_\ell\) versus \(\ell\)), at least on small angular scales (\(\lesssim 1^\circ\)), corresponding to \(\ell\gtrsim 100\). The rather precise determination of the various cosmological parameters (Planck Collaboration VI 2020) is based on this excellent agreement between theory and observation. In principle, the angular correlation function \(C(\theta)\) contains the same information as the angular power spectrum (see Eq. [\[eq:angcorr\]](#eq:angcorr){reference-type="ref" reference="eq:angcorr"}), but there are good reasons for analyzing the CMB anisotropies separately using these two approaches. First, whereas the angular power spectrum highlights the relative contributions of different spherical harmonics, the angular correlation function describes the variations of \(T\) in real space. Some features may emerge more prominently in one description instead of the other. Second, \(C(\theta)\) highlights properties of the anisotropies at relatively large angles (i.e., small \(\ell\)'s), while the power spectrum provides ample detail in the opposite regime, namely \(\ell\gg 10\). One can easily see the reason for this with a quick inspection of figure [\[f1\]](#f1){reference-type="ref" reference="f1"}. When using the full angular power spectrum, the model fit is dominated by over 1,000 values of \(\ell\) at \(\ell>30\), and while the optimized fit is striking in this regime, as we discuss elsewhere in this paper, the error bars and variance at \(\ell<30\) are much larger and the fit is significantly less compelling. On the other hand, the angular correlation function displays the temperature anisotropies with equal weighting over all angles \(0\le\theta\le 180^\circ\). When fitting a model to these data, the dominant influence is therefore at \(\theta>10^\circ\), the opposite of what happens with the power spectrum. Third, \(C(\theta)\) is a direct pixel-based measure, not requiring any reconstruction of the observed sky. Some of the systematics may therefore have less impact in one approach relative to the other. In spite of these differences, one may nonetheless reasonably expect that, of the various large-angle anomalies seen in the CMB (Planck Collaboration VII 2020), the one associated with \(C(\theta)\) could have the same physical origin as a second emerging from the angular power spectrum. Many authors (see, e.g., Copi et al. 2010) have pointed out that the most striking feature of the angular correlation function (as seen, say, in fig. 2 of Melia & López-Corredoira 2018) is not just that it disagrees with \(\Lambda\)CDM at a very high level of confidence (\(\gtrsim 3\sigma\); Copi et al. 2015), but that it is nearly zero at large angles (\(\gtrsim 60^\circ\))---in significant tension with inflation, which should have produced correlations throughout the visible Universe. Most of the statistical weight in the *Planck* analysis of the angular power spectrum is provided by the high-\(\ell\) multipoles, and one may see in figure [\[f1\]](#f1){reference-type="ref" reference="f1"} that the best-fitting portion of the curve traces the data at \(\ell\gtrsim 60\) much better than those at low-\(\ell\)'s. Indeed, a general lack of angular power appears within the range \(2\lesssim \ell\lesssim 30\)---especially with the quadrupole \(\ell=2\)---at a confidence level exceeding \(99\%\) (Bennett et al. 2011, Planck Collaboration VI 2020). Quite remarkably, none of the attempted fixes invoking physically motivated inputs, such as neutrino properties, the number of relativistic degrees of freedom, or a running of the spectral index in the fluctuation distribution, have provided any significant improvement to the fit (see, e.g., Planck Collaboration XXIV 2014; Planck Collaboration VI 2020). To be sure, the \(C(\theta)\) anomaly is much more obvious than the low harmonic-space quadrupole and octopole power (see also O'Dwyer et al. 2004) but, given that they both reference CMB anisotropic structure at very large angles (\(\gtrsim 60^\circ\)), it is tempting to consider the possibility that these two anomalies may have a common physical origin (see discussion in Planck Collaboration VII 2020). This is still not generally accepted, however. The low power at large angles may simply be due to cosmic variance (Efstathiou 2003b), in which our Universe may represent a random extrapolation away from its most likely configuration. Another possible explanation for the vanishing \(C(\theta)\) at large angles is that a *range* of low multipoles are 'conspiring' to cause the cancellation (Copi et al. 2009). In this interpretation, the combined contribution of \(C_2\)--\(C_5\) is canceled by the contributions of \(C_\ell\) with \(\ell>5\), an apparent conspiracy that seems to violate the independence of different multipole powers expected in statistical isotropy. But all such attempts at reconciling the missing angular correlations with low power at \(\ell\lesssim 30\) have thus far been inconclusive. Sentiment has therefore shifted to the idea that finding a theoretical explanation for the missing power at low \(\ell\)'s may be best served by first explaining the vanishing \(C(\theta)\) at large angles.
# Best-fit truncated spectrum
## The angular correlation function
We recently addressed this issue (Melia & López-Corredoira 2018) by attempting to show that the absence of large-angle correlations in the *Planck* data is best explained by the introduction of a non-zero minimum wavenumber \(k_{\rm min}\) in the primordial fluctuation power spectrum, \[{\mathcal{P}}(k) = \begin{cases} A_s\left({k\over k_0}\right)^{n_s-1} & \mbox{if } k\ge k_{\rm min} \\ 0 & \mbox{if } k<k_{\rm min} \end{cases}\;,\label{eq:power}\] where \(n_s\) is the spectral index, measured to have the value \(0.965\pm0.004\) by *Planck* (Planck Collaboration VI 2020); \(A_s=2.196^{+0.051}_{-0.06}\times 10^{-9}\) is the spectral amplitude; and \(k_0=0.05\) Mpc\(^{-1}\) is the scalar pivot scale. Such a cutoff challenges inflation's ability to simultaneously fix the horizon problem and produce a near scale-free distribution of quantum fluctuations matching the observed CMB anisotropies (Liu & Melia 2020; Melia 2020), because the presence of a non-zero \(k_{\rm min}\) implies a specific time at which the accelerated expansion could have started, thus restricting the number of e-folds available to expand the Universe beyond the Hubble horizon. In contrast, standard inflationary \(\Lambda\)CDM has no cutoff---and hence no limit (in principle)---to the degree of inflated expansion. This is the reason, of course, why inflation predicts an angular correlation at all scales which, however, has not been observed by any of the CMB satellite missions (Wright et al. 1996; Bennett et al. 2003; Planck Collaboration VI 2020). Our earlier analysis of the angular correlation function, based on the notion that quantum fluctuations were generated in the early Universe with this power spectrum and \(k_{\rm min}\not=0\), demonstrated that the *Planck* data rule out a zero cutoff at a confidence level exceeding \(8\sigma\). An optimized fit to \(C(\theta)\) showed that the correlations measured over the *entire* sky could be reproduced remarkably well with a cutoff \[k_{\rm min}={4.34\pm0.50\over r_{\rm dec}}\;,\] where \(r_{\rm dec}\) is the comoving distance between us and redshift \(z_{\rm dec}=1080\), at which decoupling is thought to have occurred in standard \(\Lambda\)CDM. Using the latest *Planck* parameters (Planck Collaboration VI 2020), one finds \(r_{\rm dec}\approx 13,804\) Mpc, and therefore a corresponding minimum wavenumber \[k_{\rm min}=(3.14\pm 0.36)\times 10^{-4}\;{\rm Mpc}^{-1}\;\;\;({\rm ang.\; corr.\; function})\;,\label{eq:kmin}\] roughly 160 times smaller than the pivot scale \(k_0\). This previous study, however, did not complete the analysis by investigating how such a cutoff would impact the angular power spectrum---both at low and high \(\ell\)'s. Given how well inflationary \(\Lambda\)CDM reproduces the observed \(C_\ell\)'s at \(\ell\gtrsim 60\) (see fig. [\[f1\]](#f1){reference-type="ref" reference="f1"}), the question in not merely whether a non-zero \(k_{\rm min}\) could also reduce the power of the low multipoles, but whether it could do so while simultaneously preserving the excellent agreement seen between theory and observation elsewhere in this plot. Answering this question is the principal goal of this paper.
## The angular power spectrum
The idea of reducing power at large angles by modifying the potential of inflation is not new (for a comprehensive review, see Martin et al. 2013). Other approaches have included: (1) altering inflation's initial conditions (Berera et al. 1998; Contaldi et al. 2003; Boyanovsky et al. 2006; Powell & Kinney 2007; Wang & Ng 2008; Cicoli et al. 2014; Das et al. 2015; Broy et al. 2015; Liu & Melia 2020; Melia 2020); (2) a reconsideration of the Integrated Sachs-Wolfe effect (Das & Souradeep 2014); (3) the introduction of spatial curvature (Efstathiou 2003a); (4) the use of a non-trivial topology (Luminet et al. 2003); (5) geometric effects (Campanelli et al. 2006; Campanelli et al. 2007); (6) a violation of statistical anisotropies (Hajian & Souradeep 2003); the creation of primordial micro black-hole remnants (Scardigli et al. 2011); and loop quantum cosmology (Barrau et al. 2014), among several others. Two particular approaches stand out in terms of their overlap with our treatment in this paper, notably those of Iqbal et al. (2015) and Santos et al. (2018), to which we shall return in § 4 below, when we discuss our results in a broader context. The analysis we carry out here is novel for two principal reasons: (i) it is complementary to---and required for the completion of---our previous study of the angular correlation function (Melia & López-Corredoira 2018), and (ii) it is very focused on the specific modification to the primordial fluctuation spectrum exhibited in Equation ([\[eq:power\]](#eq:power){reference-type="ref" reference="eq:power"}), which is itself highly motivated by its success in resolving the missing correlations at large angles. Can a truncated spectrum \({\mathcal{P}}(k)\) with the same \(k_{\rm min}\) resolve both the \(C(\theta)\) and low-\(\ell\) anomalies? We use the publicly available code CAMB[^1] (Lewis et al. 2000) to calculate the angular power spectrum, based on a line-of-sight approach described in Seljak & Zaldarriaga (1996). The code is modified slightly to include the additional parameter (i.e., \(k_{\rm min}\)) characterizing the primordial power spectrum (Eq. [\[eq:power\]](#eq:power){reference-type="ref" reference="eq:power"}). In addition to this cutoff, the code optimizes all of the usual standard-model parameters but, as we shall see from an inspection of figures [\[f1\]](#f1){reference-type="ref" reference="f1"}--[\[f3\]](#f3){reference-type="ref" reference="f3"} and Table [1](#table1){reference-type="ref" reference="table1"}, most of them remain unchanged from their *Planck*-\(\Lambda\)CDM values. We therefore highlight here a comparison only of the six basic parameters between the two cases in which \(k_{\rm min}=0\) and \(k_{\rm min}\not=0\): baryon density, \(\Omega_{\rm b}\), cold dark matter density, \(\Omega_{\rm c}\), the Thomson scattering optical depth, \(\tau\), due to reionization, the angular size of the acoustic horizon, \(\theta_s\), the spectral index, \(n_s\), and the scalar amplitude, \(A_s\). Each of the quantities \(\Omega_i\) represents the ratio of the energy density \(\rho_i\) of species "i\" to the critical density, \(\rho_{\rm c}\equiv 3c^2 H_0^2/8\pi G\), in terms of the Hubble constant \(H_0\). Our optimization procedure is even less sensitive to the other parameters in *Planck*-\(\Lambda\)CDM. In optimizing the fit, we use the publicly available code cosmomc (Lewis & Bridle 2002) with an initial cutoff range \(0\le k_{\rm min}\le 10^{-3}\) Mpc\(^{-1}\), and a combined data set that includes: the low-\(\ell\) TT and EE and high-\(\ell\) TT, TE and EE likelihoods from Planck 2018 (Planck Collaboration VI 2020), the Dark Energy Survey (DES) Year 1 results (Abbott et al. 2018; Troxel et al. 2018), the Baryon Acoustic Oscillation (BAO) compilation from Alam et al. (2017), and the Pantheon Type Ia SN catalog (Scolnic et al. 2018). Note that the error bars shown in figures [\[f1\]](#f1){reference-type="ref" reference="f1"} and [\[f2\]](#f2){reference-type="ref" reference="f2"} include cosmic variance, approximated as Gaussian (see fig. 1 in Planck Collaboration VI 2020), and so do the likelihoods also taken from Planck Collaboration VI (2020) used with our cosmomc calculations. The calculated angular power \(C_\ell\) for *Planck*-\(\Lambda\)CDM, with \(k_{\rm min}=0\), is indicated by the black curve in figures [\[f1\]](#f1){reference-type="ref" reference="f1"} and [\[f2\]](#f2){reference-type="ref" reference="f2"}, in comparison with the *Planck* data (blue dots). Our corresponding optimized fit using \(k_{\rm min}\) as a free parameter is represented by the red solid curve. Based on this fit, we find an optimized value \[k_{\rm min}={2.82^{+1.94}_{-1.09}\over r_{\rm dec}}\;\;\;({\rm ang.\; power\; spectrum})\;,\] corresponding to a maximum fluctuation size \(\theta_{\rm max}\approx 127^\circ\) in the plane of the sky. For a redshift \(z_{\rm dec}=1080\) at decoupling, this cutoff represents a maximum fluctuation size \(\lambda_{\rm max}\sim 31\) Mpc at that redshift. The six basic parameters for this fit, and their confidence regions, are shown in figure [\[f3\]](#f3){reference-type="ref" reference="f3"} and listed in Table [1](#table1){reference-type="ref" reference="table1"}. Note that, in spite of the fact that a \(k_{\rm min}\) compatible with zero appears to be possible within \(1\sigma\) in the \(\tau-k_{\rm min}\) panel of figure [\[f3\]](#f3){reference-type="ref" reference="f3"}, this is not the case because the \(1\sigma\) and \(2\sigma\) regions are different in 1D and 2D. The distinction arises because the \(1\sigma\) and \(2\sigma\) levels in the 2D histograms are not the \(68\%\) and \(95\%\) values used for the 1D distributions. The relevant \(1\sigma\) and \(2\sigma\) levels for a 2D histogram of samples are \(1-\exp(-0.5)\rightarrow 39.3\%\) and \(1-\exp(-2)\rightarrow 86.5\%\).[^2] Using \(39.3\%\) and \(86.5\%\) of the region, the contour plots then do exclude a \(k_{\rm min}\) compatible with zero at the \(1\sigma\) (1D) level. The fact that a non-zero cutoff \(k_{\rm min}\) has minimal, if any, impact on the optimized fit at \(\ell\gtrsim 30\), is consistent with the prevailing view that the physical origin of the fluctuations is due to different physical effects above and below \(\sim 1^\circ\). It may therefore be useful to compare this optimized value of \(k_{\rm min}\) based on the whole range of \(\ell\)'s with that found when \(\ell\) is restricted to the range (\(\lesssim 30\)) thought to be dominated by the Sachs-Wolfe effect (Sachs & Wolfe 1967). The optimized parameter values for this case, using the low-\(\ell\) TT data from *Planck* 2018, combined with the DES, BAO, and Type Ia SNe data sets, are shown in figure [\[f4\]](#f4){reference-type="ref" reference="f4"}. The corresponding best-fit cutoff \(k_{\rm min}\) for this restricted range is \((3.3^{+1.7}_{-1.3})\times 10^{-4}\) Mpc\(^{-1}\), consistent with the full-\(\ell\) range value, but even closer to the cutoff found using the angular correlation function. In either case, the cutoff 'fixing' the low-power anomaly at \(\ell\lesssim 4\) is fully consistent with the value that completely accounts for the missing correlations at angles \(\gtrsim 60^\circ\), suggesting that both of these large-angle anomalies in the standard model are due to the same truncation in \({\mathcal{P}}(k)\).
[\[table1\]]{#table1 label="table1"}
# Discussion
A side-by-side comparison of the best-fit parameters in \(\Lambda\)CDM with and without a cutoff \(k_{\rm min}\) (Table [1](#table1){reference-type="ref" reference="table1"}) reveals only negligible differences between them, but the angular power spectrum for \(\ell\lesssim 4\) is improved significantly compared with the data. Thus, the excellent agreement between standard \(\Lambda\)CDM and the CMB data for \(\ell\gtrsim 60\) is completely unaffected by our introduction of a cutoff to the primordial power spectrum. Yet, based on the optimized \(k_{\rm min}\) value using the whole \(\ell\) range (and even more so using the restricted range \(\ell\lesssim 30\)), we conclude that a zero cutoff is ruled out at a confidence level of \(\sim 2.6\sigma\). On its own, this is already interesting enough for us to study its impact on slow-roll inflation (Liu & Melia 2020; Melia 2020), but when combined with our earlier measurement of \(k_{\rm min}\) based on the angular correlation function (Melia & López-Corredoira 2018), which showed that \(k_{\rm min}=0\) was ruled out at a confidence level exceeding \(8\sigma\) for those data, these two complementary indicators argue compellingly that the primordial power spectrum \({\mathcal{P}}(k)\) must have been truncated at \(k_{\rm min}\sim 3\times 10^{-4} \;{\rm Mpc}^{-1}\). These results are completely consistent with the prevailing view that the CMB anisotropies are primarily due to two distinct physical influences: the Sachs-Wolfe effect (Sachs & Wolfe 1967) at small \(\ell\)'s and acoustic oscillations for \(\ell\gtrsim 30\). The excellent fit to the data for large \(\ell\)'s (see fig. [\[f1\]](#f1){reference-type="ref" reference="f1"}) reinforces the perception that we understand fairly well the physical process associated with small-angle anisotropies, in spite of the fact that the angular correlation function has been problematic since the beginning (Hinshaw et al. 1996; Bennett et al. 2003; Planck Collaboration VI 2020). The emerging evidence for a non-zero \(k_{\rm min}\), on the other hand, speaks directly to the cosmological expansion itself. At \(\ell\lesssim 30\), we see the effects due to the metric perturbations associated with the cosmological dynamics, and therefore the direct influence of the cosmological model itself. In this range, we are probing ever closer to the onset of the hypothesized inflationary expansion, culminating with the cutoff \(k_{\rm min}\), that signals the very first mode crossing the horizon at the beginning of the quasi-de Sitter phase produced by the inflaton field (Liu & Melia 2020; Melia 2020). The idea that at least one of the large-angle anomalies may be due to features in the primordial power spectrum, \({\mathcal{P}}(k)\), has been with us for several decades (see discussion in Planck Collaboration VII 2020). For example, Shafieloo & Souradeep (2004) assumed an exponential cutoff at low \(k\)'s, and found an optimized fit of the attenuation generally consistent with the cutoff we have measured in this paper. Their ansatz for \({\mathcal{P}}(k)\) did not include an actual truncation, however, which would be required for self-consistency with our previous treatment of \(k_{\rm min}\) in Melia & López-Corredoira (2018). The impact on inflation is different in these two cases, since a sharp termination signals a precise time at which inflation would have started. An exponential cutoff is not fully consistent with the notion of horizon-crossing for the freezing of modes. In addition, this early treatment was based on the use of WMAP data (Bennett et al. 2003), whereas our analysis incorporates the higher precision *Planck* observations (Planck Collaboration VI 2020). In figure 4, we have used only the *Planck* TT data at \(\ell\le 30\), plus other low-redshift data sets, as described in the caption. For figure 3, however, we used all of the *Planck* data, including TT, TE, EE and also both the high-\(\ell\) and other low-redshift data sets. A comparison of these results shows that the optimized value of \(k_{\rm min}\) is very similar in these two cases, suggesting that the other low-redshift data sets, such as DES, BAO and Type Ia SNe, do not significantly affect the limits on the cutoff. In related work, Nicholson & Contaldi (2009) and Hazra et al. (2014) modeled a 'dip' in \({\mathcal{P}}(k)\) on a scale \(k\sim 0.002\) Mpc\(^{-1}\) for the WMAP data. Their results confirmed those of an alternative approach (Ichiki et al. 2010), in which an oscillatory modulation was identified around \(k\sim 0.009\) Mpc\(^{-1}\). These are not a truncation, however, nor are they consistent with our analysis (Melia & López-Corredoira 2018) of a cutoff \(k_{\rm min}\) in the angular correlation function based on the *Planck* measurements. Tocchini-Valentini et al. (2005) modeled both a dip at \(k\sim 0.035\) Mpc\(^{-1}\) and a 'bump' at \(k\sim 0.05\) Mpc\(^{-1}\), also using WMAP data. As with the others, however, these features are not the same as an actual truncation \(k_{\rm min}\), and our analysis utilizes the latest *Planck* data, rather than the less precise WMAP measurements. Subsequent work by these authors, notably Tocchini-Valentini et al. (2006), improved on this analysis considerably, though they were still fully reliant on WMAP. These authors found evidence for three features in \({\mathcal{P}}(k)\), one of which is analogous to our \(k_{\rm min}\), including a cutoff at \(\sim 0.0001-0.001\) Mpc\(^{-1}\), with a confidence level of about \(2\sigma\). Our analysis based on the improved *Planck* data produces a tighter constraint on this cutoff, more in line with the \(k_{\rm min}\) found from the angular correlation function. For example, their possible cutoff range exceeds the actual value we have measured for \(k_{\rm min}\) in this paper. More recently, Hunt & Sarkar (2014, 2015) found evidence in the WMAP data of a cutoff \(k < 5 \times 10^{-4}\) Mpc\(^{-1}\) to \({\mathcal{P}}(k)\), but concluded that more accurate data, such as those provided by Planck, would need to be used to confirm these results more robustly. The work most closely aligned with our analysis in this paper is that of Iqbal et al. (2015) and Santos et al. (2018). These two groups arrived at opposite conclusions to each other, however. Though each carried out a thorough analysis of the possible cause for the low CMB power at small \(\ell\)'s, and actually agreed on approximately what value a cutoff (which they called \(k_c\), analogous to our \(k_{\rm min}\)) would be needed to address this anomaly, they adopted different model-selection criteria and produced opposing outcomes. This state of uncertainty has been with us for many years. The approach of Iqbal et al. (2015) and Santos et al. (2018) was to focus solely on the low-\(\ell\) anomaly and to compare 7 or 8 different models, using a range of model-selection tools. Our approach in this paper has been diametrically opposite to this. We have only one goal in mind, motivated by the physics of inflation. Most importantly, our previous analysis demonstrated that a zero value for \(k_{\rm min}\) is ruled out at over \(8\sigma\). Notice, e.g., how different this outcome is compared to that of Santos et al. (2018), who concluded that standard inflationary cosmology, without any adjustment to \({\mathcal{P}}(k)\), is preferred by the CMB data and their choice of model selection criteria. The work we have carried out in this paper is based on the conclusion of our previous work suggesting that only a simple cutoff \(k_{\rm min}\) in \({\mathcal{P}}(k)\) is well motivated by the CMB angular correlation function. We have not compared different models, nor have we based our analysis on a subjective choice of model selection criteria, which seem to produce the divergent outcomes seen, e.g., in Iqbal et al. (2015) and Santos et al. (2018). We showed in Liu & Melia (2020) that the cutoff \(k_{\rm min}\) measured via the CMB angular correlation function apparently rules out all inflationary models based on slow-roll potentials, including those that may have been preceded by radiation-dominated and/or kinetic-dominated phases (see also Handley et al. 2014). With such a cutoff, inflation could not simultaneously have solved the horizon problem and accounted for the observed CMB fluctuation spectrum. This is the reason why a confirmation of \(k_{\rm min}\) based on the angular power spectrum is so essential. The two large-angle anomalies are not necessarily related (see also Planck Collaboration VII 2020), so our previous result with \(k_{\rm min}\) does not by itself ensure that it is also consistent with the much better studied angular power spectrum, nor that it can also mitigate the missing power at low-\(\ell\)'s. This paper has therefore been highly focused on these two issues alone. Nevertheless, the fact that our measured value of \(k_{\rm min}\) is consistent with that found by Iqbal et al. (2015) and Santos et al. (2018) is crucial because it demonstrates a consistent outcome by three different groups. We mention in passing that an extension of the work reported by Iqbal et al. (2015) and Santos et al. (2018) to include the angular correlation function would be very beneficial. Our previous assessment was that only a simple cutoff in \({\mathcal{P}}(k)\) could completely account for the missing correlations at large angles. It would be interesting to see if an independent examination of these data, using a range of other models with various model selection tools, such as those explored by these two groups, could alter this conclusion. In this paper, we have met our primary goal of showing that the introduction of a non-zero cutoff in \({\mathcal{P}}(k)\) does not at all affect the excellent agreement between the predictions of \(\Lambda\)CDM and the CMB angular power spectrum at \(\ell\gtrsim 30\). This is critical because it is widely believed that the origin of the CMB anisotropies in this domain is well understood, given that it relies on well-established astrophysical principles. We have also demonstrated that the origin of the missing power in the low-\(\ell\) multipoles, particularly at \(\ell=2\)--\(5\), is very likely due to the same physical influence impacting the missing correlations at large angles. The two optimized values of \(k_{\rm min}\), i.e., \((3.14\pm0.36)\times 10^{-4}\) Mpc\(^{-1}\) for the latter, and \((2.04^{+1.4}_{-0.79})\times 10^{-4}\) Mpc\(^{-1}\) for the former, are fully consistent with each other. These complementary results reinforce the view that we are beginning to see direct evidence of the initiation of inflation, if this process actually did occur. Nevertheless, the disparity between the primordial fluctuation spectrum expected under these conditions and what is actually required to produce the CMB anisotropies, increases the tension between the concordance model and the observations.
# Conclusion
The existence of large-angle anomalies in the temperature is uncontested though, given the modest significance with which they disagree with the standard model, and the fact that they were detected a posteriori, makes it unclear how much evidence they actually provide for a true, physical origin (Planck Collaboration VII 2020). They may simply be statistical fluctuations, even though such an outcome appears to be highly unlikely. Nevertheless, if any of them do indeed correspond to a real physical effect, it would be extremely important for us to confirm this, justifying the continued attention paid to these unusual features. As noted earlier, however, we must still temper our conclusions regarding the reality of these large-angle anomalies, given that the Planck Collaboration VII (2020) study found only weaker evidence for them in the polarization data. This is why upcoming missions designed specifically to measure the CMB polarization will be so critical to the continuation of this work. For example, LiteBIRD has been selected for development and launch in 2028 by the Japan Aerospace Exploration Agency, with the goal of mapping the CMB polarization over the whole sky with unprecedented precision (Hazumi et al. 2019). Its required angular coverage will correspond to \(2\le\ell\le 200\), perfectly attuned to the requirements for the analysis in this paper. Perhaps even more impressively, the Probe of Inflation and Cosmic Origins (PICO) mission (Hanany et al. 2019) is still in the study phase, but is projected to be an imaging polarimeter scanning the sky for 5 years in 21 frequency bands spread between 21 and 799 GHz. If selected, PICO will produce a full-sky survey of the intensity and polarization of the CMB with a final combined-map noise level equivalent to an amazing 3,300 Planck missions. And as a prominent third example, the Cosmic Origins Explorer (CORE) mission (Delabrouille et al. 2018) proposed to the European Space Agency is projected to have 19 frequency channels spanning the \(60-600\) GHz range, and an angular resolution from \(2^\prime\) to \(18^\prime\). It will observe the entire sky repeatedly over four years of continuous scanning. Its design will be optimized for complementarity with ground-based observations, performing observations---such as that required by the analysis carried out in this paper---essential to CMB polarization science not achievable without a dedicated space mission. Proposed, or in-progress, experiments to map the CMB polarization from the ground include the Simons Observatory (Ade et al. 2019), located in the high Atacama Desert in Northern Chile inside the Chajnator Science Preserve, at an altitude of 5,200 meters. Though its goals are currently not as lofty as those of the space-based missions, its principal aim is to produce a polarization map of the sky with an order of magnitude better sensitivity than *Planck*. The principal aim of all these future missions is, of course, to detect the B-mode polarization in the tell-tale signature of gravity waves generated during inflation. Nevertheless, a more precise measurement of the E-mode polarization anisotropies at a much higher sensitivity than that available to *Planck* should answer the question of whether or not the large-angle anomalies seen in the temperature are also confirmed in the polarization maps. This would then provide compelling evidence that they are not merely due to secondary, background or instrumental effects, further motivating a study of the impact on inflationary theory of a cutoff, \(k_{\rm min}\), in the primordial power spectrum, \({\mathcal{P}}(k)\). For example, as shown by Liu & Melia (2020), a \(k_{\min}\) like that measured earlier in our analysis of the angular correlation function (Melia & López-Corredoira 2018), and reinforced by the work reported in this paper, which is based exclusively on the angular power spectrum, rules out the majority---if not all---of the slow-roll inflaton potentials proposed thus far. If one insists on inflation simultaneously fixing the horizon problem and accounting for the observed primordial power spectrum, \({\mathcal{P}}(k)\), then the accelerated expansion resulting from these hypothesized fields misses the required comoving distance by a factor \(\sim 10\). Moreover, neither a radiation-dominated, nor a kinetic-dominated, phase preceding inflation can alleviate this disparity (Liu & Melia 2020; see also Handley et al. 2014). Our results reinforce the growing view that, at a minimum, inflation probably needs to be modified or, at worst, needs to be replaced, in order to conform with these observations.
[^1]: www.camb.info
[^2]: https://corner.readthedocs.io/en/latest/pages/sigmas.html. | {'timestamp': '2021-09-14T02:21:14', 'yymm': '2109', 'arxiv_id': '2109.05480', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05480'} |
# Introduction {#sec:1}
Black hole thermodynamics is one of the most interesting recent discoveries of theoretical physics. Bekenstein argued that the area \(A\) of a black hole horizon has the properties of the entropy \(\mathcal{S}\) in ordinary thermodynamics and must be proportional to it (the proportionality factor was determined later). However, the similarity between black hole area and thermodynamical entropy did not make sense initially because it was believed that black holes are cold objects, until Hawking discovered that the Schwarzschild black hole emits quantum radiation with a blackbody spectrum at temperature \(T_\mathrm{H}= \frac{1}{8\pi G M}\) (now called the Hawking temperature), where \(M\) is the Schwarzschild mass of the black hole. Here we use geometrized units in which the speed of light \(c\), the Boltzmann constant \(K_B\), and the reduced Planck constant \(\hbar\) are unity. Hawking's discovery of black hole radiation originates from the application of quantum field theory to curved spacetime and, by extension, it implies that all black holes must radiate, making them thermal objects and completing Bekenstein's suggestion. The Bekenstein-Hawking entropy \(\mathcal{S}= A/4\) and the Hawking temperature \(T_\mathrm{H}\) allow for the construction of a self-consistent black hole thermodynamics (, see for reviews), which is now an important part of modern theoretical physics. Adding a cosmological constant, as in the Schwarzschild-de Sitter/Kottler and Schwarzschild-Anti-de Sitter black holes, adds richness to the thermodynamical behaviour of black holes. In addition to creating multiple horizons which could be viewed as thermodynamical sub-systems, it leads to the possibility of the Hawking-Page phase transition , which was later interpreted in the context of the AdS/CFT correspondence as the counterpart of the deconfinement transitions for the conformal field theory living on the Anti-de Sitter boundary . The negative cosmological constant turns out to play the role of the pressure, to be added to the thermodynamical picture, thus extending the structure of the phase space and making phase transitions possible. In this context, a rich literature on "black hole chemistry" has emerged and black hole thermodynamics has taken a new lease on life (*e.g.*, [@Kubiznak:2014zwa; @Karch:2015rpa; @Kubiznak:2015bya; @Kubiznak:2016qmn; @Sinamuli:2017rhp; @Tzikas:2018cvs; @Astefanesei:2019ehu; @Mir:2019ecg; @Kumar:2020cve] and references therein). Originally, it came as a surprise that the Bekenstein-Hawking entropy is proportional to the black hole area and not to its volume, as in ordinary thermodynamics where entropy is an extensive quantity proportional to the mass, and then to the volume, of a system. This feature largely remains a mystery. Recent literature has discussed the possibility of replacing the Bekenstein-Hawking entropy with other entropy notions based on non-extensive statistics [@Bialas:2008fa; @Tsallis:2012js; @Huang:2014pda; @Brustein:2014iha; @Nishioka:2014mwa; @Czinner:2015eyk; @Dong:2016fnf; @Wen:2016itq; @Czinner:2017tjq; @Qolibikloo:2018wqw; @Johnson:2018bma; @Tannukij:2020njz; @Promsiri:2020jga; @Samart:2020klx; @Ren:2020djc; @Mejrhit:2020dpo; @Nakarachinda:2021jxd; @Abreu:2021avp], such as the Rényi and Tsallis entropies. However, changing the entropy notion is risky because entropy enters many thermodynamical equations and other quantities need to be modified in order to keep the whole construction of thermodynamics self-consistent. While modifying entropy is challenging and many authors have focused on this task, here we point out the risks inherent in these modifications for related thermodynamics. In particular, it is problematic to modify the Hawking temperature of blackbody radiation based only on non-extensive statistics, while there are independent arguments pointing toward the "correct" choice of thermodynamical energy and black hole mass. When all these aspects are considered together, it turns out to be quite difficult to replace the Bekenstein-Hawking entropy based solely on the idea of non-extensive statistics while keeping black hole thermodynamics self-consistent. In this work we follow the notation of Ref. : the metric signature is \({-}{+}{+}{+}\) and the units are such that the speed of light \(c\), the Boltzmann constant \(K_B\), and the reduced Planck constant \(\hbar\) are unity.
# Standard black hole thermodynamics {#sec:2}
The geometry of the Schwarzschild black hole is described by the line element \[\label{dS3BB} ds^2=-f(r) dt^2 + \frac{dr^2}{f(r)} + r^2 d\Omega^2_{(2)}\,, \quad\quad f(r) \equiv 1-\frac{2GM}{r} \,,\] where \(G\) is Newton's constant, \(M\) is the black hole mass, and \(d\Omega^2_{(2)}=d\vartheta^2 +\sin^2 \vartheta \, d\varphi^2\) is the line element on the unit two-sphere. The black hole event horizon is located at the Schwarzschild radius \[\label{horizonradius} r_\mathrm{H}=2GM\,.\] By considering quantum field theory on the spacetime with this horizon, Hawking discovered that the Schwarzschild black hole radiates with a blackbody spectrum at the temperature \[\label{dS6BB} T_\mathrm{H} = \frac{1}{8\pi GM}\.\] The Hawking temperature can be also understood geometrically. When \(r\sim r_\mathrm{H}\), we define \(\delta r\) by \(r \equiv r_\mathrm{H} + \delta r\). Then by Wick-rotating the time coordinate \(t\to i\tau\), the line element ([\[dS3BB\]](#dS3BB){reference-type="ref" reference="dS3BB"}) is recast as \[\label{TH1} ds^2 \simeq \frac{\delta r}{r_\mathrm{H}} \, d\tau^2 + \frac{r_\mathrm{H}}{\delta r} \, d(\delta r)^2 + r_\mathrm{H}^2 \, d\Omega^2_{(2)}\,.\] We further define a new radial coordinate \(\rho\) by \(d\rho = d \left( \delta r\right) \sqrt{ r_\mathrm{H}/ \delta r}\), that is \[\label{TH2} \rho = 2 \sqrt{ r_\mathrm{H} \delta r} \quad\quad \mbox{or} \quad\quad \delta r= \frac{\rho^2}{4 r_\mathrm{H}} \,,\] in terms of which the line element ([\[TH1\]](#TH1){reference-type="ref" reference="TH1"}) becomes \[\label{TH3} ds^2 \simeq \frac{\rho^2}{ 4 r_\mathrm{H}^2} \, \rho^2d\tau^2 + d\rho^2 + r_\mathrm{H}^2 \, d\Omega_{(2)}^2 \,.\] In order to avoid the conical singularity in the Wick-rotated Euclidean space around \(\rho\sim 0\), we require the periodicity of the Euclidean time coordinate \(\tau\) \[\label{TH4} \frac{\tau}{ 2 r_\mathrm{H}} \sim \frac{\tau}{ 2 r_\mathrm{H}} + 2\pi \.\] Because the inverse of the period \(t_0\) of the Euclidean time coordinate corresponds to the temperature, as in the Euclidean path integral formulation of the finite temperature field theory for any field \(\phi\), \[\label{TH5} \int \left[ D\phi \right] \e^{ \, \int_0^{t_0} dt \, L(\phi)} = \mathrm{Tr}\left( \, \e^{-t_0 H} \right) = \mathrm{Tr}\left( \, \e^{-\frac{H}{T} } \right) \,,\] one finds that the Schwarzschild black hole has temperature \(T\), which is nothing but the Hawking temperature ([\[dS6BB\]](#dS6BB){reference-type="ref" reference="dS6BB"}), \[\label{TH6} T = \frac{1}{4\pi r_\mathrm{H}}= \frac{1}{8\pi GM} \equiv T_\mathrm{H} \.\] Thus, the Hawking temperature can be obtained solely from the geometry of the spacetime endowed with the event horizon. We can derive also the entropy \(\mathcal{S}\) from the geometrical point of view. Eq. ([\[TH5\]](#TH5){reference-type="ref" reference="TH5"}) tells us that the partition function \(Z(T)\) and the free energy \(F(T)\) are given by \[\label{FZ} \e^{-\frac{F(T)}{T}} = Z(T) = \mathrm{Tr}\left(\, \e^{-\frac{H}{T} }\right) = \int \left[ D\phi \right] \e^{S(\phi)} \,,\] with the periodic boundary condition that the Euclidean time has period \(1/T\). In Eq. ([\[FZ\]](#FZ){reference-type="ref" reference="FZ"}), \(S(\phi)=\int_0^{t_0} dt \, L(\phi)\) is the Euclidean action. In the low-temperature regime when \(T\) is sufficiently small, the path integral ([\[FZ\]](#FZ){reference-type="ref" reference="FZ"}) can be estimated as \[\label{FZ2} \int \left[ D\phi \right] \e^{S(\phi)} \sim \e^{S \left( \phi_\mathrm{cl} \right)}\] in the WKB approximation, where \(\phi_\mathrm{cl}\) is a classical solution of the field equations given by the Euclidean action \(S(\phi)\). Then Eq. ([\[FZ\]](#FZ){reference-type="ref" reference="FZ"}) allows us to estimate the free energy \(F\) which, in turn, gives the entropy \(\mathcal{S}\) by using the thermodynamical relations. In order to estimate the free energy \(F\), we consider the Schwarzschild-Anti-de Sitter geometry \[\begin{aligned}
\label{SAdS1} && ds^2 =-f(r) dt^2 + \frac{dr^2}{f(r)} + r^2 d\Omega_{(2)}^2\,,\\ &&\nonumber\\ && f(r) \equiv 1-\frac{2GM}{r} + \frac{r^2}{l^2}\,, \end{aligned}\] where \(\Lambda=-3/l^2\) is the cosmological constant (the reason why we consider the Schwarzschild-Anti-de Sitter spacetime instead of the Schwarzschild one with the flat Minkowski background will be explained later). We now rewrite the function \(f(r)\) in the form \[\begin{aligned}
&& f(r) = \frac{\left(r-r_\mathrm{H} \right) \left( r + \frac{r_\mathrm{H}}{2} + i a \right) \left( r + \frac{r_\mathrm{H}}{2}-i a \right)}{l^2 r} \,,\nonumber\\ &&\\ && r_\mathrm{H}\left( \frac{r_\mathrm{H}^2}{4} + a^2 \right) = 2GM l^2 \,, \quad -\frac{3r_\mathrm{H}^2}{4} + a^2 = l^2 \.\nonumber\\ && \label{SAdS2} \end{aligned}\] Eliminating \(a\) in the last two equation yields \[\label{SAdSC1} r_\mathrm{H} \left( r_\mathrm{H}^2 + l^2 \right) = 2GM l^2 \,.\] When \(r\sim r_\mathrm{H}\), \(f(r)\) behaves as \[\label{SAdS4} f(r) \sim \frac{\frac{ 9r_\mathrm{H}^2}{4}+a^2}{l^2 r_\mathrm{H}} \left(r-r_\mathrm{H} \right)\] and the Hawking temperature is \[\label{SAdS5} T_\mathrm{H} \simeq \frac{\frac{ 9r_\mathrm{H}^2}{4}+a^2}{4 \pi l^2 r_\mathrm{H}} = \frac{3 r_\mathrm{H}^2 + l^2}{4 \pi l^2 r_\mathrm{H}} = \frac{1}{4\pi l^2}\left( 3 r_\mathrm{H} + \frac{l^2}{r_\mathrm{H}} \right) \,,\] where we have used the last equation ([\[SAdS2\]](#SAdS2){reference-type="ref" reference="SAdS2"}) to substitute for \(a\). After the Wick rotation, the action becomes \[\begin{aligned}
S &=& \frac{1}{16\pi G} \int d^4 x \sqrt{-g} \left( R + \frac{6}{l^2} \right) \nonumber\\ &=&-\frac{3}{2Gl^2} \int_0^{ 1/ T_\mathrm{H} } dt \int_{r_\mathrm{H}}^L dr \, r^2 \nonumber\\ &=& \frac{r_\mathrm{H}^3-L^3}{2G\, l^2 T_\mathrm{H}} \,,\label{SAdS3} \end{aligned}\] where we have introduced a cutoff \(L\) to regulate the divergence of the action ([\[SAdS3\]](#SAdS3){reference-type="ref" reference="SAdS3"}). It is instructive to consider the difference between the action of the Schwarzschild-Anti-de Sitter spacetime and that of the pure Anti-de Sitter spacetime. We determine the period of the Euclidean time \(1/\tilde T_\mathrm{H}\) in Anti-de Sitter space so that the physical length (between \(t=0\) and \(t=1/{\tilde T}_\mathrm{H}\)) equals the physical length (between \(t=0\) and \(t= 1/ T_\mathrm{H}\)) in the Schwarzschild-Anti-de Sitter spacetime: \[\label{SAdSB1} \left( 1-\frac{2GM}{L} + \frac{L^2}{l^2} \right)^{1/2} \frac{1}{T_\mathrm{H}} = \left( 1 + \frac{L^2}{l^2} \right)^{1/2} \frac{1}{{\tilde T}_\mathrm{H}} \,,\] or \[\label{SAdSB2} \frac{1}{{\tilde T}_\mathrm{H}} = \frac{1}{T_\mathrm{H}} \sqrt{ 1-\frac{2GM}{L\left( 1 + \frac{L^2}{l^2} \right)}} \sim \frac{1}{T_\mathrm{H}} \left( 1-\frac{GMl^2}{L^3} \right) \,.\] Then, the action \(S_\mathrm{AdS}\) for Anti-de Sitter spacetime is \[\begin{aligned}
S_\mathrm{AdS} &=&-\frac{3}{2Gl^2} \int_0^{1/{\tilde T}_\mathrm{H} } dt \int_0^L dr \, r^2 =-\frac{L^3}{2G l^2\, {\tilde T}_\mathrm{H}} \nonumber\\ & \sim &-\frac{L^3}{2G l^2 T_\mathrm{H}} + \frac{GMl^2}{2G l^2 T_\mathrm{H}} \nonumber\\ &=&-\frac{L^3}{2G l^2 T_\mathrm{H}} + \frac{r_\mathrm{h} \left( r_\mathrm{H}^2 + l^2 \right) }{4G l^2 T_\mathrm{H}} \,,\label{SAdS3B} \end{aligned}\] where we have used Eq. ([\[SAdSC1\]](#SAdSC1){reference-type="ref" reference="SAdSC1"}). In the limit \(L\to \infty\), the action \(S_\mathrm{BH}\) for the black hole in Anti-de Sitter space reduces to \[\label{BH1} S_\mathrm{BH} = S-S_\mathrm{AdS} = \frac{r_\mathrm{H} \left( r_\mathrm{H}^2-l^2 \right) }{4G \, l^2 T_\mathrm{H}}\] and the free energy \(F\) is \[\label{SAdS8} F=-T_\mathrm{H} S_\mathrm{BH} =-\frac{r_\mathrm{H} \left( r_\mathrm{H}^2-l^2 \right) }{4G l^2} \,.\] Upon use of the thermodynamical relations \[\label{SAdS9} E = F-T_\mathrm{H}\, \frac{d F}{d T_\mathrm{H}}\,, \quad \quad \mathcal{S} = \frac{E-F}{T_\mathrm{H}} =-\frac{d F}{d T_\mathrm{H}}\] for the thermodynamical energy \(E\) and the entropy \(\mathcal{S}\), we find the entropy \[\begin{aligned}
\label{SAdS10} \mathcal{S} =&-\frac{ dF/d r_\mathrm{H} }{ d T_\mathrm{H}/ d r_\mathrm{H}} = \frac{ \frac{3 r_\mathrm{H}^2-l^2 }{4G l^2}}{ \frac{1}{4\pi l^2}\left( 3-\frac{l^2}{r_\mathrm{H}^2} \right) } = \frac{\pi r_\mathrm{H}^2}{G}=\frac{A}{4G}\,, \end{aligned}\] where \(A = 4\pi {r_\mathrm{H}}^2\) is the horizon area. The expression ([\[SAdS10\]](#SAdS10){reference-type="ref" reference="SAdS10"}) is valid in the limit of a flat Minkowski background \(l^2\to \infty\), thus we obtain the Bekenstein-Hawking entropy from a geometrical viewpoint. We should note, however, that the action ([\[SAdS3\]](#SAdS3){reference-type="ref" reference="SAdS3"}) vanishes in the Minkowski background \(l^2\to \infty\), hence, the "Schwarzschild" black hole in Anti-de Sitter space with finite \(l\) somehow plays the role of the regularization of the Schwarzschild black hole in flat Minkowski background. Another consideration is in order: restoring the constants, the Hawking temperature and the Bekenstein-Hawking entropy read T\_= , = , then the free energy \(F=E-TS\) does not contain the reduced Planck constant \(\hbar\) and, in this sense, it is a classical quantity like \(E\). Non-extensive black hole entropies that are inspired by modified uncertainly principles or quantum gravity corrections should not affect \(F\) unless they correct \(T_\mathrm{H}\). Then, according to Eq. ([\[SAdS10\]](#SAdS10){reference-type="ref" reference="SAdS10"}), the entropy \(\mathcal{S}\) is only affected through corrections to the black hole temperature. However, if this temperature remains the Hawking temperature \(T_\mathrm{H}\), \(\mathcal{S}\) cannot receive corrections, or else also Eq. ([\[SAdS10\]](#SAdS10){reference-type="ref" reference="SAdS10"}) must be modified accordingly. Although a bit hand-waving,[^1] this argument makes the point that thermodynamics (including black hole thermodynamics) must be a consistent theory and modifying this or that thermodynamic quantity *ad hoc* usually has unwanted consequences for the rest of the theory.
# Impossibility of non-area law entropy {#sec:3}
In the previous section, we have summarized the derivation of both the Hawking temperature and the Bekenstein-Hawking entropy from the viewpoint of the geometry. As is well known, the area law for the Bekenstein-Hawking entropy can always be obtained if we identify the thermodynamical energy \(E\) with the black hole mass \(M\), \(E=M\), and the temperature of the system with the Hawking temperature ([\[dS6BB\]](#dS6BB){reference-type="ref" reference="dS6BB"}) , \(T=T_\mathrm{H}\). In fact, the thermodynamical relation \(dE= T d\mathcal{S}\) yields \[\label{S1} d\mathcal{S}=\frac{dE}{T}=8\pi GM dM =d\left( 4 \pi G M^2 \right) \,,\] that is, \[\label{S2} \mathcal{S}=4 \pi G M^2 + \mathcal{S}_0\,,\] where \(S_0\) is an integration constant. If we assume \(\mathcal{S}=0\) when \(M=0\), that is, in the absence of the black hole, then \(\mathcal{S}_0=0\) and \[\label{S3} \mathcal{S}= \frac{\pi {r_\mathrm{H}}^2}{G} = \frac{A}{4G}\] where \(A\equiv 4\pi {r_\mathrm{H}}^2\) is the horizon area. Thus, we have shown that the Bekenstein-Hawking entropy (*i.e.*, the area law for the black hole entropy) can be obtained by assuming \(E=M\) and \(T=T_\mathrm{H}\) by using the thermodynamical relation \(d\mathcal{S}=dE/T\). Two questions arise naturally:
1. Can we identify the thermodynamical energy \(E\) with the black hole mass \(M\) (*i.e.*, \(E=M\))? Furthermore, if the black hole is not Schwarzschild nor isolated, there is no Arnowitt-Deser-Misner mass: should then \(M\) be the quasilocal mass contained in the horizon sphere, or the "black hole part" of it? If so, which quasilocal mass? Several quasilocal mass prescriptions exist in the literature (see Ref. for a review).
2. Is the temperature of the black hole given by the Hawking temperature, \(T=T_\mathrm{H}\)?
To answer the first question, consider the following *gedankenexperiment*: assume that there is an infalling spherically symmetric shell of dust with mass \(M\) and initial radius sufficiently large. By virtue of the Birkhoff theorem , the spacetime outside the shell is the Schwarzschild one ([\[dS3BB\]](#dS3BB){reference-type="ref" reference="dS3BB"}), where \(M\) in ([\[dS3BB\]](#dS3BB){reference-type="ref" reference="dS3BB"}) is the mass of the shell. Inside the shell, spacetime is empty and flat. The shell collapses, its radius becoming smaller and smaller. When the shell crosses its Schwarzschild radius ([\[horizonradius\]](#horizonradius){reference-type="ref" reference="horizonradius"}), a black hole is formed. The geometry is always asymptotically flat and the shell mass \(M\) appearing in the Schwarzschild line element is surely the energy \(E\) of the system, \(E=M\). This energy is conserved during the collapse of the shell because, due to the Birkhoff theorem, the geometry outside of it does not change during the collapse: the energy of the final black hole must be the mass of the shell.[^2] Therefore, the thermodynamical energy \(E\) is the black hole mass \(E=M\). Regarding the mass concept used, the answer is easy for the Schwarzschild black hole: the Schwarzschild mass appearing in the line element is the obvious choice. It coincides with the Misner-Sharp-Hernandez quasilocal mass \(M_\mathrm{MSH}\) defined in any spherically symmetric spacetime by - = \^c R \_c R , where \(R\) is the areal radius (which coincides with \(r\) in the Schwarzschild case ([\[dS3BB\]](#dS3BB){reference-type="ref" reference="dS3BB"})). Other quasilocal masses do not reproduce the standard Hawking temperature and the Bekenstein-Hawking entropy. For example, consider the Brown-York quasilocal energy as a possible candidate to the role of thermodynamical energy. For a spherically symmetric metric of the form ds\^2 =-N\^2(t,r) dt\^2 + +r\^2 d\_(2)\^2 , the Brown-York energy is E\_= ̇For the Schwarzschild metric, the Brown-York energy is radius-dependent: E\_= (1- ) ; at the Schwarzschild horizon \(r=r_\mathrm{H}\), we have \(E_\mathrm{BY}=2M\). If we identify the thermodynamical energy with the Brown-York energy and the black hole temperature with the Hawking temperature, the relation \(Td\mathcal{S}=dE\) yields d\_= = 16G M dM and, integrating, \_= 8M\^2 = , which is unphysical. The only way to reconcile the Brown-York energy prescription with the relation \(Td\mathcal{S}=dE\) is by introducing the Brown-York temperature \(T_\mathrm{BY}=2T_\mathrm{H}\), which disagrees with Hawking's fundamental result. One does not see how (or why) Hawking's derivation of the black hole temperature should be modified to make it agree with the Brown-York mass prescription in thermodynamics. Let us consider now the Schwarzschild-Anti-de Sitter black hole (parallel considerations apply to the Schwarzschild-de Sitter case). For the Schwarzschild-Anti-de Sitter black hole, the Misner-Sharp-Hernandez mass is[^3] M\_ = M- , where it is easy to isolate the black hole contribution from the one due to the negative cosmological constant, and the Hawking temperature and the Bekenstein-Hawking entropy should refer to these contributions (at least in the limit of small black holes). This happens also for more general black holes embedded in a cosmological "background" described by the time-dependent McVittie metric (this geometry contains the Schwarschild-de Sitter/Kottler metric as a special case) [@Carrera:2009ve; @Faraoni:2015ula]. However, this is not the case for other cosmological black holes, for example for the Sultana-Dyer black hole for which the Misner-Sharp-Hernandez/Hawking-Hayward mass includes a third term coupling the black hole and the cosmological energies. However it can be argued that, at least for the Sultana solution, the time-dependence of both black hole and cosmological horizons precludes a discussion with equilibrium thermodynamics while, in general, an adiabatic expansion for dynamical black holes should recover the Schwarzschild thermodynamics. It is interesting that the thermodynamics of dynamical apparent horizons (as opposed to static null event horizons) employs the Hawking-Hayward quasilocal energy as the thermodynamical energy of a black hole, which reduces to the Misner-Sharp-Hernandez mass in spherical symmetry. Due to the time-dependence, the first law of thermodynamics must be generalized to include an energy supply vector . In spherical symmetry, the Clausius definition of entropy coincides with the entropy obtained from Wald's Noether charge method [@Wald:1993nt; @Iyer:1994ys] supplemented by the Kodama flow, and the Kodama temperature (which reduces to \(T_\mathrm{H}\) in the static case) . Now, both of these prescriptions give back the Bekenstein-Hawking entropy in the static case, which receives support from the more involved discussion of dynamical black hole horizons. Another indication of the privileged role of the Misner-Sharp-Hernandez mass among the spectrum of quasilocal masses available in the literature comes from the fact that it is the Noether charge associated with the conservation of the Kodama current (which always holds in spherical symmetry ). It is the Misner-Sharp-Hernandez mass that is used as thermodynamical energy in the thermodynamics of dynamical apparent horizons: this is a self-consistent thermodynamical picture that uses in an essential way the Kodama time, in the absence of a timelike Killing vector (see Ref. for a review of the relevant tunneling formalism). Replacing the black hole mass used in the thermodynamics of static black hole event horizons with another energy (quasilocal or not) would imply the same replacement in the thermodynamics of apparent horizons and would make it inconsistent. Regarding the second question about the temperature, note that the Hawking radiation is obtained if the geometry with horizon is prescribed and the standard Hawking temperature is the parameter appearing in the thermal distribution of the emitted Hawking radiation. If we place the black hole in a heat bath at temperature \(T\), thermal equilibrium between the black hole radiation and the heat bath occurs when the radiation temperature equals the temperature of the heat bath, \(T=T_\mathrm{H}\). Therefore, we can use the heat bath as a thermometer. The temperature measured by the heat bath must be the standard Hawking temperature of the Hawking radiation and, therefore, we identify the latter with the black hole temperature. Entropy constructs other than the Bekenstein-Hawking one are often considered in the literature, for example the Rényi entropy \[\label{RS1} \mathcal{S}_\mathrm{R}=\frac{1}{\alpha} \ln \left( 1 + \alpha \, \mathcal{S} \right) \,,\] where \(\alpha\) is a parameter and \(\mathcal{S}\) is the Bekenstein-Hawking entropy ([\[S3\]](#S3){reference-type="ref" reference="S3"}), which is recovered in the limit \(\alpha\rightarrow 0\). In this case, Eq. ([\[S2\]](#S2){reference-type="ref" reference="S2"}) with \(\mathcal{S}_0=0\) gives \[\label{RS2} \mathcal{S}_\mathrm{R}=\frac{1}{\alpha} \ln \left( 1 + 4 \pi \alpha G M^2 \right) \,.\] If the mass \(M\) coincides with the energy \(E\) of the system due to the energy conservation as in, in order for this system to be consistent with the thermodynamical equation \(d\mathcal{S}=dE/T\), one needs to define the "Rényi temperature" \(T_\mathrm{R}\) by \[\label{TR1} \frac{1}{T_\mathrm{R}} \equiv \frac{d\mathcal{S}_\mathrm{R}}{dM} = \frac{8 \pi G M}{1 + 4 \pi \alpha G M^2} \,,\] that is, \[\label{TR2} T_\mathrm{R} = \frac{1}{8\pi GM} + \frac{\alpha M}{2} = T_\mathrm{H} + \frac{\alpha}{16 \pi G T_\mathrm{H}}\] which is, of course, quite different from the Hawking temperature \(T_\mathrm{H}\). Therefore, this "Rényi temperature" \(T_\mathrm{R}\) is not the temperature perceived by any observer detecting Hawking radiation, which is the concept that started the research area of black hole thermodynamics. This fact tells us that the "Rényi temperature" \(T_\mathrm{R}\) is likely physically irrelevant for black hole thermodynamics. In Eq. ([\[TR1\]](#TR1){reference-type="ref" reference="TR1"}), we assumed that the thermodynamical energy \(E\) is the black hole mass \(M\) and we obtained an unphysical result. One wonders what the result would be if we assumed that the thermodynamical temperature \(T\) coincides with the Hawking temperature \(T_\mathrm{H}\), instead of assuming \(E=M\). Under the assumptions \(T=T_\mathrm{H}\) and \(\mathcal{S}=\mathcal{S}_\mathrm{R}\), using \(dE=Td\mathcal{S}\) we find the corresponding thermodynamical energy \(E_\mathrm{R}\): \[\begin{aligned}
dE_\mathrm{R} &=& T_\mathrm{H} d\mathcal{S}_\mathrm{R} = \frac{1}{8\pi GM} \, \frac{8 \pi GM dM}{1 + 4\pi \alpha G M^2}\\ &&\nonumber\\ &=& \frac{dM}{1 + 4\pi \alpha G M^2} \label{RE1} \end{aligned}\] and, integrating, \[\label{RE2} E_\mathrm{R} = \frac{\arctan \left( \sqrt{4\pi \alpha \, G}\, M \right) }{ \sqrt{4\pi \alpha \, G}} = M-\frac{4\pi \alpha \, GM^2 }{3} + \mathcal{O}\left( \alpha^2 \right) \,,\] where we have fixed the integration constant so that \(E_\mathrm{R}=0\) when \(M=0\). Due to the correction \(-\frac{4\pi \alpha GM^2 }{3} + \mathcal{O}\left( \alpha^2 \right)\), the expression ([\[RE2\]](#RE2){reference-type="ref" reference="RE2"}) of the thermodynamical energy \(E_\mathrm{R}\) obtained differs from the black hole mass \(M\), \(E_\mathrm{R}\neq E\), in a way that has no obvious physical interpretation. More important, if the spherically symmetric dust shell collapses to a Schwarzschild black hole, the result obtained above seems to conflict with energy conservation. Another entropy notion encountered frequently in the literature and motivated by non-extensive statistics is the Tsallis entropy , which suggests as a potential alternative to the Bekenstein-Hawking entropy for black holes the quantity \[\label{TS1} \mathcal{S}_\mathrm{T} = \frac{A_0}{4G} \left( \frac{A}{A_0} \right)^\delta\] instead of the Rényi-like entropy ([\[RS1\]](#RS1){reference-type="ref" reference="RS1"}) (, see also ). Here \(A_0\) is a constant with the dimensions of a length squared and \(\delta\) is a parameter that quantifies the non-extensivity. In the case \(\delta = 1\), we obtain the standard Bekenstein-Hawking entropy ([\[SAdS10\]](#SAdS10){reference-type="ref" reference="SAdS10"}) or ([\[S3\]](#S3){reference-type="ref" reference="S3"}). In Ref. , cosmology with the Tsallis-type entropy ([\[TS1\]](#TS1){reference-type="ref" reference="TS1"}) has been studied. The corresponding late-time universe contains an effective dark energy, which could be phantom or quintessence, without an effective cosmological constant. One obtains an effective cosmological constant from the generalized non-extensive Tsallis type entropy also in the inflationary era of the early universe. As done in Eq. ([\[TR1\]](#TR1){reference-type="ref" reference="TR1"}), assuming that the thermodynamical energy \(E\) is given by the black hole mass \(M\), we find \(A=4\pi \left( 2GM \right)^2 = 16\pi G^2 E^2\) and the corresponding Tsallis entropy \[\label{TS2} \mathcal{S}_\mathrm{T} = \frac{A_0}{4G} \left( \frac{16\pi G^2 E^2}{A_0} \right)^\delta \,.\] We may define also the "Tsallis temperature" \[\begin{aligned}
T_\mathrm{T} & \equiv & \frac{dE}{d\mathcal{S}_\mathrm{T}} = \frac{2G}{\delta A_0 E^{2\delta-1}} \left( \frac{A_0}{16\pi G^2} \right)^\delta \nonumber\\ &=& \frac{2G}{\delta A_0 M^{2\delta-1}} \left( \frac{A_0}{16\pi G^2} \right)^\delta \label{TS3} \end{aligned}\] which is, of course, different from the Hawking temperature ([\[TH6\]](#TH6){reference-type="ref" reference="TH6"}) unless \(\delta =1\). Then, instead of identifying the black hole mass \(M\) with the thermodynamical energy \(E\), we could assume that the temperature is the Hawking temperature ([\[TH6\]](#TH6){reference-type="ref" reference="TH6"}). Then, since \(A= 4\pi \left( 4\pi T_\mathrm{H} \right)^{-2} = \frac{1}{4\pi {T_\mathrm{H}}^2}\), we would find \[\label{TS4} \mathcal{S}_\mathrm{T} = \frac{{A_0}^{1-\delta}}{4 G \left(4\pi {T_\mathrm{H}}^2 \right)^\delta}\] and then we may define the "Tsallis energy" \(E_\mathrm{T}\) by \[\label{TS5} d E_\mathrm{T} = T_\mathrm{H} d\mathcal{S}_\mathrm{T} =-\frac{\delta {A_0}^{1-\delta} d T_\mathrm{H}}{2 G \left(4\pi\right)^\delta {T_\mathrm{H}}^{2\delta}}\] or, integrating, \[\label{TS6} E_\mathrm{T} = \frac{\delta {A_0}^{1-\delta} }{2 \left( 2\delta-1 \right) G \left(4\pi\right)^\delta {T_\mathrm{H}}^{2\delta-1}} = \frac{\delta {A_0}^{1-\delta} \left( 8\pi GM \right)^{2\delta-1} }{2 \left( 2\delta-1 \right) G \left(4\pi\right)^\delta } \,,\] where we have again fixed the integration constant by imposing \(E_\mathrm{T}=0\) at \(M=0\). The standard relation \(E_\mathrm{T} = M\) is reproduced for \(\delta=1\), but \(E_\mathrm{T} \neq M\) otherwise. In order to build a thermodynamic theory one needs to identify, in addition to the entropy, the thermodynamical energy \(E\) and the temperature \(T\). If we choose as entropy any quantity different from the Bekenstein-Hawking entropy, then we lack physical grounds and we can choose the thermodynamical energy \(E\) and/or the temperature \(T\) as we wish. Not surprisingly, this procedure leads to an unphysical outcome, *i.e.*, it conflicts with energy conservation and/or with the established physics of the Hawking radiation process. To conclude, the assumed Rényi entropy ([\[RS1\]](#RS1){reference-type="ref" reference="RS1"}) cannot be the black hole entropy, although it might be the entropy of a system different from the black hole (see, for example, Ref. ), as in the case of the Tsallis entropy. The Tsallis entropy is obtained as a result of the Fermi, Bose, or Boltzmann statistics for the system with long-range forces by the standard statistical mechanics procedure of using the Hamiltonian and counting the number of states. In this sense, there is no physical counterpart that could be the Rényi statistics. The Rényi entropy could be an index specifying the information, with no relation with the statistics of any physical system.
# Discussion and conclusions {#sec:4}
The idea of replacing the standard Boltzmann-Gibbs statistics with non-extensive statistics has led to the Rényi and Tsallis entropies. Since the Bekenstein-Hawking black hole entropy is non-extensive, several authors have considered the possibility of replacing the Bekenstein-Hawking entropy with the Rényi or the Tsallis one. This proposal has somehow mixed with the parallel idea of correcting the Bekenstein-Hawking entropy with modifications due to the Generalized Uncertainty Principle, or with other ideas. We have pointed out the fact that changing the entropy goes hand in hand with changing other thermodynamical quantities, or else the entire thermodynamical theory may become inconsistent, but this creates more problems of principle. For spherical black holes, one could think of reconciling the Hawking temperature \(T_\mathrm{H}\) with some entropy notion different from the Bekenstein-Hawking entropy by adopting as thermodynamical energy some quasilocal energy \(E_x\), for which various prescriptions can be found in the literature. We have discussed the Brown-York energy as an example. However, already at first sight, this task appears very difficult to say the least (and, even if it was logically possible, very contrived from the physical point of view). The explicit forms of the Rényi and Tsallis entropies make it practically impossible to match \(T_\mathrm{H}\), \(E_x\), and \(\mathcal{S}_\mathrm{R}\) (or \(\mathcal{S}_\mathrm{T}\)) while satisfying the relation \(Td\mathcal{S}_\mathrm{R,T}= dE_x\) even for the simple Schwarzschild black hole, which is static and isolated. Indeed, the case for the Bekenstein-Hawking entropy seems rather compelling at this point. Let us examine now possible loopholes to the area law for entropy. In the previous sections, we have shown that the area law of the Bekenstein-Hawking entropy can always be obtained if we assume that the thermodynamical energy \(E\) is identified with the mass \(M\), \(E=M\), and the temperature of the system is the Hawking temperature. Moreover, other forms of the entropy such as the Rényi and the Tsallis entropies lead to unphysical results. In order to show that \(E=M\), we have used the argument of the dust shell collapse in conjunction with energy conservation. This argument might not carry over to theories of modified gravity, especially higher derivative theories such as \(F(R)\) gravity. In the Einstein gravity, the falling shell of the dust constitutes an exact solution of the field equations. In higher derivative gravity, the junction conditions between the matter and the vacuum at the shell becomes more complex and more restrictive [@Barrabes:1997kk; @Bressange:2000qp; @Velay-Vitow:2017odc; @Berezin:2019maf; @Berezin:2020mas; @Berezin:2020bvs; @Aviles:2019xae; @Olmo:2020fri; @Chu:2021uec; @Rosa:2021teg]. Therefore, whether the infalling dust shell constitutes an exact solution of the field equations remains to be determined and the previous argument may not apply. What, is more the gravitational constant \(G\) becomes a scalar degree of freedom \(\phi \sim G^{-1}\), the Brans-Dicke-like scalar field, already in scalar-tensor gravity (which includes \(F(R)\) theories as a subclass ). Then, the area law for the entropy becomes \(\mathcal{S}= \phi A/4\) and the scalar \(\phi\) gives a contribution to the differential \(d\mathcal{S}\) in \(Td\mathcal{S}=dE\). While it is true that all vacuum black holes of scalar-tensor gravity that are spherical, asymptotically flat, static, and sit in a minimum of the potential \(V(\phi)\) for \(\phi\) reduce to the Schwarzschild black hole [@Hawking:1972qk; @Bekenstein:1996pn; @Sotiriou:2011dz; @Bhattacharya:2015iha], the discussion becomes more complicated already for asymptotically de Sitter black holes. Even in Einstein gravity, if we include quantum corrections there might be modifications to the reasoning above. For example, quantum fluctuations of the horizon have been discussed in. The quantum fluctuation may effectively increase the area of the horizon, changing the entropy. In the arguments following Eq. ([\[FZ2\]](#FZ2){reference-type="ref" reference="FZ2"}) we have used the WKB approximation, which is valid at low temperature. But at high temperature the correction given by the \(1/T_\mathrm{H}\) expansion becomes large and there could be a violation of the entropy area law. The \(1/T_\mathrm{H}\) correction may be also regarded as a quantum corrections because if we include \(\hbar\), the factor \(t_0= 1/T_\mathrm{H}\) in front of the Hamiltonian \(H\) in ([\[TH5\]](#TH5){reference-type="ref" reference="TH5"}) becomes \(\frac{t_0}{\hbar} =\frac{1}{\hbar T_\mathrm{H}}\). Such corrections are considered in . Then, if we assume \(E=M\) as in ([\[TR1\]](#TR1){reference-type="ref" reference="TR1"}), the Hawking temperature might not be modified and the thermal distribution of the radiation might be changed from that of blackbody radiation. Such a correction might appear as the shift from \(\alpha \rightarrow 0\) in the Rényi-type entropy ([\[RS1\]](#RS1){reference-type="ref" reference="RS1"}) or the shift from \(\delta=1\) in the Tsallis-like entropy ([\[TS1\]](#TS1){reference-type="ref" reference="TS1"}). In general, entropy corresponds to the number of physical degrees of freedom of a thermodynamical system. The renormalization of a quantum theory implies that the number of degrees of freedom depends on the scale. In standard field theory massive modes decouple in the low-energy regime and, therefore, the number of degrees of freedom decreases. The situation is more complicated in the case of gravity: if gravity is described by string theory, an infinite tower of massive modes appears at high temperature, which introduces the upper bound to the temperature called Hagedorn temperature. Even from a naive point of view, if the spacetime fluctuations become large in the high temperature regime, the number of degrees of freedom may increase. If, instead, gravity becomes topological, the number of degrees of freedom will decrease, which could be consistent with holography. In the case of the Bekenstein-Hawking entropy ([\[S3\]](#S3){reference-type="ref" reference="S3"}) or ([\[SAdS10\]](#SAdS10){reference-type="ref" reference="SAdS10"}), because the area is given by \(A= \frac{1}{4\pi {T_\mathrm{H}}^2}\), in the high temperature regime of large \(T_\mathrm{H}\), the entropy decreases. This happens, of course, because the smaller black hole has higher Hawking temperature. The decrease in the entropy corresponds to the loss of physical degrees of freedom due to Hawking radiation. In the case of the Rényi-type entropy ([\[RS1\]](#RS1){reference-type="ref" reference="RS1"}), if the parameter \(\alpha\) is positive, the deviation from the Bekenstein-Hawking entropy \(\mathcal{S}\) ([\[S3\]](#S3){reference-type="ref" reference="S3"}) becomes large in the low-temperature region where \(\mathcal{S}\) becomes large, which might contradict the above speculations in which this deviation becomes large in the high-temperature region instead. If \(\alpha<0\), we may further modify the Rényi-like entropy as \[\label{RS1M1} \mathcal{S}_\mathrm{R}=\frac{1}{\alpha} \ln \left| 1 + \alpha \mathcal{S} \right| \,;\] then there is a singularity when \(1 + \alpha \mathcal{S} =0\), or \[\label{RS1M2} T_\mathrm{H} = T_\mathrm{c} \equiv \sqrt{-\frac{\alpha}{16\pi G}} \,.\] If we start at temperature higher than \(T_\mathrm{c}\), the critical temperature \(T_\mathrm{c}\) gives a lower bound on the Hawking temperature. If, instead, we start at temperature lower than \(T_\mathrm{c}\), then \(T_\mathrm{c}\) sets an upper bound that might correspond to the Hagedorn temperature. Therefore, it might be interesting to consider a model in which \(\alpha\) depends on the temperature and becomes negative at high temperature. In the case of the Tsallis-like entropy ([\[TS1\]](#TS1){reference-type="ref" reference="TS1"}), if the parameter \(\delta\) is larger than unity, \(\delta > 1\), the Tsallis entropy ([\[TS1\]](#TS1){reference-type="ref" reference="TS1"}) becomes smaller than the Bekenstein-Hawking entropy at high temperature. If \(\delta < 1\), the Tsallis type entropy becomes larger than the Bekenstein-Hawking entropy at high temperatures. Then, it might be interesting to consider a model in which the parameter \(\delta\) depends on the temperature. If \(\delta>1\) at high temperatures, this model might correspond to the model becoming topological at high \(T\). If, instead, \(\delta<1\) in the high-temperature regime, the model might correspond to a violent fluctuation of spacetime or to string theory. If the Rényi entropy ([\[RS1\]](#RS1){reference-type="ref" reference="RS1"}) comes from a quantum correction, we may express its parameter as \(\alpha=\alpha_0 \hbar\) and expand Eq. ([\[RS1\]](#RS1){reference-type="ref" reference="RS1"}) with respect to \(\hbar\) as \[\label{RS1hbar} \mathcal{S}_\mathrm{R}=\frac{1}{\alpha_0 \hbar} \ln \left( 1 + \alpha_0 \hbar \mathcal{S} \right) = \mathcal{S}-\frac{\alpha_0 \hbar}{2} \, \mathcal{S}^2 + \frac{\alpha_0^2 \hbar^2}{3} \, \mathcal{S}^3 + \cdots\] If \(\alpha\) is positive, the leading correction is negative, which could hint at gravity becoming topological by quantum effects, whereas if \(\alpha<0\) the leading correction is positive, possibly saying that the number of degrees of freedom increases as in string theory. Similarly, by writing \(\delta = 1 + \delta_0 \hbar\), we may expand the Tsallis entropy ([\[TS1\]](#TS1){reference-type="ref" reference="TS1"}) as \[\begin{aligned}
\mathcal{S}_\mathrm{T} &=& \frac{A_0}{4G} \left( \frac{A}{A_0} \right)^{1 + \delta_0 \hbar} \nonumber\\ &=& \frac{A}{4G} \left[ 1 + \delta_0 \hbar \ln \frac{A}{A_0} + \frac{\delta_0^2 \hbar^2}{2!} \left( \ln \frac{A}{A_0} \right)^2 + \cdots \right] \. \nonumber\\ && \label{TS1hbar} \end{aligned}\] Then, if \(\delta>0\) the number of degrees of freedom increases for large black holes where \(A>A_0\) and decreases for small black holes with \(A<A_0\). If \(\delta<0\), the number of degrees of freedom increases for small black holes and decreases for large ones. We should also note that the logarithmic correction in Eq. ([\[TS1hbar\]](#TS1hbar){reference-type="ref" reference="TS1hbar"}) often appears at the first loop in quantum field theory. Another possible loophole might be the case in which there are two horizons, as discussed in Ref. , where the entropies of the Reissner-Nordström and of the Kerr black holes have been investigated, and it is claimed that there might be contributions to the entropy from the correlation of the two horizons. As another example, we may consider the Schwarzschild-de Sitter/Kottler spacetime, where two horizons (the black hole and the cosmological horizon) appear. The line element is \[\begin{aligned}
&&ds^2=-f(r) dt^2 + \frac{dr^2}{f(r)} + r^2 d\Omega_{(2)}^2\, ,\\ && f(r) = 1-\frac{2GM}{r}-\frac{r^2}{l^2}\,; \label{SdS1} \end{aligned}\] if we assume \(l^2>0\), we can rewrite \(f(r)\) as \[\begin{aligned}
&& f(r) =-\frac{\left(r-r_-\right) \left( r-r_+ \right) \left( r + r_+ + r_-\right)}{l^2 r} \, ,\label{SdS2a}\\ &&\nonumber\\ && r_+ r_-\left( r_+ + r_-\right) = 2GM l^2 \,,\label{SdS2b}\\ &&\nonumber\\ && r_+^2 + r_-^2 + r_+ r_-=l^2 \,,\label{SdS2c} \end{aligned}\] where we assume \(r_+>r_-\). Then, \(r_+\) and \(r_-\) are the radii of the cosmological and black hole horizons, respectively. When \(r\sim r_-\), \(f(r)\) behaves as \[f(r) \sim \frac{\left( r_+-r_-\right) \left( 2 r_-+ r_+ \right)}{l^2 r_-} \left(r-r_-\right) \, ,\label{SdS4}\] and the Hawking temperature \(T_\mathrm{H}^\mathrm{(bh)}\) of the black hole horizon is \[\begin{aligned}
T_\mathrm{H}^\mathrm{(bh)} &=& \frac{\left( r_+-r_-\right) \left( 2 r_- + r_+ \right)}{4\pi l^2 r_-} \nonumber\\ &=& \frac{1}{4\pi l^2}\frac{-2 r_-^2 + r_-r_+ + r_+^2}{r_-} = \frac{1}{4\pi l^2}\frac{ l^2-3 r_-^2 }{r_-} \nonumber\\ &=& \frac{1}{4\pi l^2}\left(\frac{ l^2}{r_-}-3 r_-\right) \. \label{SdS5} \end{aligned}\] Similarly, the Gibbons-Hawking temperature of the cosmological horizon is \[\label{SdS5B} T_\mathrm{H}^\mathrm{(c)} = \frac{\left( r_+-r_-\right) \left( r_-+ 2 r_+ \right)}{4\pi l^2 r_+}\] and we should note that \[\label{SdS5C} T_\mathrm{H}^\mathrm{(bh)}-T_\mathrm{H}^\mathrm{(c)} = \frac{ \left( r_+-r_-\right)^2 \left( r_+ + r_-\right) }{4\pi l^2 r_+ r_-} \geq 0 \,,\] which implies that there could be heat flow from the black hole to the cosmological horizon. In Eq. ([\[SdS5C\]](#SdS5C){reference-type="ref" reference="SdS5C"}), the equality "\(=\)" holds if and only if \(r_+=r_-\), that is, in the extremal case corresponding to the Nariai spacetime. Because \(T_\mathrm{H}^\mathrm{(bh)} \neq T_\mathrm{H}^\mathrm{(c)}\) in general, even if we Wick-rotate the spacetime into the Euclidean signature, we cannot remove both conical singularities corresponding to the black hole and the cosmological horizons and the calculation of the free energy performed for the Schwarzschild-Anti-de Sitter spacetime could not be applied. In the following, we consider the case in which only the conical singularity of the black hole horizon is removed by assuming the period of the Euclidean time \(1/T_\mathrm{H}^\mathrm{(bh)}\), but the conical singularity of the cosmological horizon remains. The equations ([\[SdS2b\]](#SdS2b){reference-type="ref" reference="SdS2b"}), ([\[SdS2c\]](#SdS2c){reference-type="ref" reference="SdS2c"}) tell us that \[\begin{aligned}
\label{SdS6} && r_-\left( 2 r_+ + r_-\right) \frac{dr_+}{dM} + r_+ \left( r_+ + 2 r_- \right) \frac{dr_-}{dM} = 2G l^2 \, ,\nonumber\\ &&\\ && \left( 2 r_+ + r_-\right) \frac{dr_+}{dM} + \left( r_+ + 2 r_-\right) \frac{dr_-}{dM} = 0 \,, \end{aligned}\] which give \[\begin{aligned}
\frac{dr_+}{dM} &=&-\frac{2Gl^2}{\left( r_--r_+ \right) \left( 2r_+ + r_-\right)} \, ,\\ &&\nonumber\\ \frac{dr_-}{dM} &=&-\frac{2Gl^2}{\left( r_+-r_-\right) \left( r_+ + 2 r_-\right)} =-\frac{2Gl^2}{r_+^2 + r_+ r_--2 r_-^2 }\nonumber\\ &&\nonumber\\ &=&-\frac{2Gl^2}{l^2-3 r_-^2 } \,. \label{SdS7} \end{aligned}\] Then the action is \[\begin{aligned}
\label{SdS3} S=&\frac{1}{16\pi G} \int d^4 x \sqrt{-g} \left( R-\frac{6}{l^2} \right) \\ &\nonumber\\ = & \frac{3}{2Gl^2} \int_0^{ 1/ T_\mathrm{H}^\mathrm{(bh)} } dt \int_{r_-}^{r^+} dr \, r^2 \nn &\nonumber\\ = & \frac{r_+^3-r_-^3}{2G l^2 T_\mathrm{H}} = \frac{ \left( r_+-r_-\right) \left( r_+^2 + r_-^2 + r_+ r_- \right)}{2G l^2 T_\mathrm{H}^\mathrm{bh}} = \frac{ r_+-r_-}{2GT_\mathrm{H}^\mathrm{bh}} \end{aligned}\] and the free energy \(F\) reads \[\label{SdS8} F= T_\mathrm{H}^\mathrm{(bh)} S = \frac{r_+-r_-}{2G} \,.\] By using the thermodynamical relations for the thermodynamical energy \(E\) and the entropy \(\mathcal{S}\) in ([\[SAdS8\]](#SAdS8){reference-type="ref" reference="SAdS8"}), we then find \[\begin{aligned}
\label{SdS10} \mathcal{S} =&-\frac{1}{2G} \frac{ -\frac{2Gl^2}{\left( r_--r_+ \right) \left( 2r_+ + r_-\right)} + \frac{2Gl^2}{\left( r_+-r_-\right) \left( r_+ + 2 r_-\right)} } {-\frac{1}{4\pi l^2}\left(-\frac{ l^2}{r_-^2}-3 \right) \frac{2Gl^2}{l^2-3 r_-^2 }} \nn &\nonumber\\ = &-\frac{2 \pi l^2 r_-^2}{G} \frac{ 2 r_-+ r_+ + 2 r_+ + r_-}{\left( r_+-r_-\right) \left( r_+ + 2 r_-\right) \left( 2r_+ + r_-\right)} \frac{ l^2-3 r_-^2}{l^2 + 3 r_-^2} \nn &\nonumber\\ = &-\frac{6 \pi l^2 r_-^2}{G} \frac{r_+ + r_-}{\left( 2r_+ + r_- \right) \left(l^2 + 3 r_-^2 \right)}\,, \end{aligned}\] which is negative and, as a result, the quantity \(\mathcal{S}\) given by Eq. ([\[SdS10\]](#SdS10){reference-type="ref" reference="SdS10"}) cannot be identified with the entropy. We cannot obtain the area law entropy for the Schwarzschild-de Sitter spacetime, which might be due to the correlation between the two horizons. A possible way to avoid the above problem of the Schwarzschild-de Sitter spacetime could be to analytically continue the results ([\[SAdS8\]](#SAdS8){reference-type="ref" reference="SAdS8"}) and ([\[SAdS10\]](#SAdS10){reference-type="ref" reference="SAdS10"}) to the Schwarzschild-Anti-de Sitter spacetime by replacing \(l^2\to-l^2\), obtaining \[\label{SdS8B} F=-\frac{r_\mathrm{H} \left( {r_-}^2 + l^2 \right) }{4G l^2} \,, \quad \quad \mathcal{S} = \frac{-\frac{3 {r_-}^2 + l^2 }{4G l^2}}{-\frac{1}{4\pi l^2}\left( 3 + \frac{l^2}{{r_-}^2} \right) } =\frac{A}{4G}\,,\] which could be the standard result. The differences between Eqs. ([\[SdS10\]](#SdS10){reference-type="ref" reference="SdS10"}) and ([\[SdS8B\]](#SdS8B){reference-type="ref" reference="SdS8B"}) originate from the integration region of the action, as we find by comparing Eqs. ([\[SAdS3\]](#SAdS3){reference-type="ref" reference="SAdS3"}) and ([\[SdS3\]](#SdS3){reference-type="ref" reference="SdS3"}) and also come from the subtraction of the background in ([\[BH1\]](#BH1){reference-type="ref" reference="BH1"}). This fact tells us that, if we can use the analytic continuation \(l^2\to-l^2\), the domain of integration should include the region outside the cosmological horizon \(L\gg r_+\), which could induce a conceptual problem in black hole physics. Anyway, the existence of two horizons makes it difficult to judge what could be the correct prescription to evaluate the entropy. In, the holographic dark energy model based on the Tsallis-type entropy or its generalization associated with the cosmological horizon was proposed. We may consider the creation of a black hole in the holographic dark energy. If the holographic dark energy has a higher temperature than the black hole temperature, there could occur an energy flow to the black hole and the black hole may consequently grow, which may signal an instability of the holographic dark energy. Black hole seeds may be generated by fluctuations of the energy density, which could be the density of the holographic dark energy. If we regard the holographic dark energy as a kind of effective perfect fluid, however, it should have a negative effective pressure. This negative pressure generates a repulsion in the fluid, which protects a fluctuation from growing and becoming a black hole. In this sense, the collapse of the holographic dark energy to black holes could be avoided and the holographic dark energy could be stable. The flow of heat energy onto a black hole, if it occurs due to the temperature difference between black hole and surrounding dark energy, is difficult to model. Ref. models the spherical accretion of phantom dark energy onto a black hole in the test fluid approximation but, realistically, the backreaction is not described and the same will be true for radial heat flow. It may be possible to describe the net heat transferred in an event occurring between two stationary phases, but this would probably require a cosmological evolution that is completely *ad hoc.* To conclude, the entropy area law is physically well motivated, unlike its Rényi and Tsallis potential competitors, and solid motivations seem to be needed before departures from it can be taken too seriously. In particular, replacing the Bekenstein-Hawking entropy with the Rényi, Tsallis, or another entropy is premature and rather arbitrary at the current stage of knowledge. | {'timestamp': '2021-10-05T02:34:58', 'yymm': '2109', 'arxiv_id': '2109.05315', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05315'} |
null | null |
null | null |
# Introduction
River deposits on Mars record past climatic conditions and thus serve as a proxy for Mars' climate evolution towards its current cold and dry state. Since the end of the era of regionally-integrated valley networks in the early Hesperian \(\sim3.4\) Ga, fluvial activity on Mars primarily produced shorter landforms, including small, 'pollywog', exit breach craters, alluvial fans, and gullies (which may or may not be fluvial in origin; see ). Some of the alluvial fans are associated with aqueous minerals (hydrated silica; e.g. ). The apparently recent ages of these fluvial features are at odds with the belief that intense atmospheric loss early in Mars' history would have rapidly rendered Mars' surface cold and dry. Further, Amazonian/Hesperian catastrophic floods are commonly (but not exclusively) believed to require a \(\geq1\) km thick cryosphere (and thereby a cold climate) in order to over-pressurize the aquifers that fed these floods. Multiple mechanisms have been proposed to explain late-forming rivers, including impact-induced snowmelt/precipitation, transient greenhouse atmospheres, and obliquity-shift induced snowmelt/precipitation. Each of these scenarios predicts transient excursions from an otherwise cold, dry climate state. Different proposed mechanisms have different dependencies on volcanism and/or \(p\)CO\(_2\), both of which declined over time. Thus, the timing of young fluvial activity is an important constraint on Martian climate evolution. In this study, we focus on the alluvial fans (for the purposes of this study, we also include deltas, which are fans formed in standing bodies of water) and how improved statistical methodologies shed light on their formation ages. The alluvial fans are of particular interest because many are precipitation-fed, at least some formed over an extended time span (\(> 20-200\) Myr), and some are up to \(\sim 1\) km thick, recording a minimum of 100 years -- 1 Myr of intermittent river flow. These observations, in addition to the widespread distribution of alluvial fans within observed latitude bands, are consistent with a synoptic water source rather than localized impact trigger and therefore provide true constraints on the Martian climate through time. The primary method for estimating the absolute age of geologic surfaces on Mars is through counting the craters in an area and comparing the observed density to a chronology function, which relates crater density to absolute age. Dating small areas like fan surfaces requires counting small craters which are, by virtue of their size, more susceptible to survey complications like atmospheric filtering and geologic modification (e.g. erosion, burial; ). Indeed, derived ages for small areas such as individual deltas vary over a huge time range. To reduce these uncertainties, previous workers have averaged over numerous close-proximity fans. However, this methodology (1) implicitly assumes that fan formation occurred simultaneously for each fan (such that their surfaces reflect a single age), (2) requires that there has not been sufficient, potentially spatially-varying, resurfacing of the fan surfaces to obscure any craters in the considered size range, and (3) sets aside the likely possibility that alluvial fans formed intermittently over tens or hundreds of Myr, rather than instantaneously. We propose that these shortcomings can be vaulted by a new approach involving careful consideration of the fans' cross-cutting relationships and choice of statistical modeling methodology. The majority of alluvial fans are hosted on the interior walls of impact craters, where erosional backwasting of crater rims provided alluvium that was transported down-slope and deposited on crater floors. These fans unambiguously post-date their host-craters, whereas determination of alluvial fan stratigraphic relationships within host crater interiors requires more detailed investigation on an individual fan basis. Further, the ages of the impact craters can be estimated via crater counts on their ejecta deposits (where preserved), which provide us order-of-magnitude greater count areas (and thus the ability to count larger craters) than the fan surfaces themselves. In the following sections, we demonstrate how combining a global database of impact craters, a global geologic map, a global database of alluvial fans, and a novel statistical model can constrain alluvial fan formation timing to have persisted until \(< 3\) Ga. Further, we will discuss limitations of our approach and how updated datasets will affect fan-timing inferences.
# Datasets and Methods
To investigate the formation of alluvial fans, we used a recently compiled global database of alluvial fans (including potential deltas) in craters that incorporates previously published surveys and abstracts in addition to newly identified fans using images from the Context Camera (CTX). Although the use of CTX imagery enabled surveying smaller fans than previous studies, we expect that the smallest features are missing from the survey, due to the finite resolution of the images used. For this study, we restricted the fans dataset to those contained within the 212 Amazonian-Hesperian Impact Unit (AHi) subunits from the global geologic map of Mars that are fully contained within \(40^{\circ}\) latitude of the equator (Appendix Figure 1). These subunits are impact craters that have well-preserved rims and ejecta.These AHi subunits provide us a sample of areas that are globally distributed spatially, share similar steep slopes and topography, and record their ages via crater-counts on their ejecta. Because fans on the crater walls must post-date the crater formation, ages obtained from the ejecta provide robust upper bounds on each fan's age. Further, not all AHi subunits possess fans (70 out of 212 subunits contain fans), which enables comparison between the formation times of fan-bearing and non-fan-bearing surfaces. To obtain crater densities for each AHi subunits's ejecta, we used a global database of impact craters \(> 1\) km in diameter. However, due to concerns over survey completeness for small craters (Stuart Robbins, personal communication), we restricted the dataset to craters with diameters of at least 4 km. Manual inspection of craters on the AHi subunits revealed three additional potential data quality issues. First, ejecta blankets often contain easily-identified chains of secondary craters. Second, some craters mapped on ejecta actually pre-date ejecta, as evidenced by cross-cutting relationships with the central crater's rim, impact ejecta, or secondary craters (Figure [\[fig:fig1\]](#fig:fig1){reference-type="ref" reference="fig:fig1"}). Finally, craters in the interior of AHi subunits' central craters are often obscured from surveys due to resurfacing by burial/erosion. To remedy these crater-counting issues, we manually inspected each crater in Thermal Emission Imaging System (THEMIS) and CTX imagery and removed obvious secondary craters (elongated craters often in chains extending from the center of large nearby impact craters) and craters that obviously pre-date crater ejecta based on cross-cutting relationships. In general, we were conservative and did not discard ambiguous craters, preventing us from erroneously under-estimating the age of the subunits. This procedure, along with removing craters within each AHi subunits's central crater's interior, reduced our total crater count from 1768 to 1308. Further, we restricted our count areas on each AHi subunit to the annulus extending from the crater rim to 2\(\times\) the subunit-forming-crater's radius (i.e., we restricted our count areas to 1-2 crater radii from the crater center). In the \(< 5\%\) of cases where the AHi subunit is defined by the ejecta blankets of 2 large craters, we chose the larger crater as our nominal center. In many cases, the ejecta blankets do not extend to 2\(\times\) the crater radius, in which case we included in our count area only the intersection of the ejecta blanket and our defined annulus. We imposed the outer limit on our count area to accommodate the fact that craters "poking through" (i.e. pre-dating ejecta) near the ejecta edges cause crater densities to increase by a factor of \(\sim2\)--\(3\) but that unambiguous cross-cutting relationships are increasingly difficult to identify as the ejecta thins radially outward. This indicates that the fraction of craters "poking-through" the ejecta blankets increases near the edges, and thus, the outer limit prevents erroneously biasing our AHi subunit ages upward. Imposing the outer annulus limit further reduced our total crater count to 498.
# Statistical Modeling and Results
Our goal was to take crater counts on the AHi subunits' ejecta blankets, note which subunits host alluvial fans, and make quantitative statements about the timing and duration of alluvial fan formation. We achieved this in two steps. First, we developed a probabilistic understanding of the age of each AHi subunit by estimating their posterior age distributions in a Bayesian framework (similar to @MICHAEL2016). Second, we implemented two statistical models that represent hypothetical formation scenarios for the alluvial fans. Parameter estimation for these models enabled inferences about the timing of alluvial fan formation.
## Dating the AHi subunits
described the use of Bayesian posteriors for estimating planetary surface ages, under the name "Poisson Timing Analysis." In this framework, crater counts were assumed to be generated as a Poisson process, with underlying parameter likelihoods, \(\mathcal{L}\), given by the standard equation: \[\mathcal{L}(t|k,A) = P(k|t,A) = \frac{(A\lambda_t)^k}{k!}e^{-A\lambda_t}\] where \(k\) is the observed number of craters, \(A\) is the area of the count region, and \(\lambda_t\) is the expected density of craters on a surface with age \(t\). The correspondence between age and expected density is known as the chronology function, and throughout this study, we adopt the chronology function of (based on and ): \[\lambda_t(D) = C(D)[(3.79 \times 10^{-14})(e^{6.93t}-1) + (5.84 \times 10^{-4})t]\] where \(\lambda_t(D)\) is the expected density of craters \(> D\) km in diameter, and \(C(D)\) is a function that accounts for the size-frequency distribution of craters (Table 1 in ). If, a priori, each candidate age is assumed to be equally likely, the likelihood function can be normalized (scaled such that its integral equals 1) and treated as a posterior probability distribution for subunit age. From there, Michael et al. (2016) computed median ages, as well as the \(25\%\) and \(75\%\) age percentiles. These corresponded to their final age estimates and error bounds, respectively. In our framework, we had strong reason to believe a non-uniform prior distribution in AHi subunit formation age. In particular, impact fluxes on Mars were elevated \(> 3\) Ga, relative to today (see among many others, ). This implies that the AHi subunits, which are themselves formed by impact craters, are more likely to have formed early in Mars' history. Thus, for our application, we took the prior cumulative distribution function (CDF) of AHi subunit ages to be proportional to the chronology function, truncated at \(4.5\) Ga. A sensitivity test truncating at \(3.54\) Ga is discussed in §4.). We then computed the prior probability density function (PDF) on AHi subunit ages as the normalized derivative of the chronology function (Figure [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"}): \[P(t) \propto \frac{d}{dt} [(3.79 \times 10^{-14})(e^{6.93t}-1) + (5.84 \times 10^{-4})t].\] Using the likelihood function and prior distribution, we computed posterior age PDF's (Figure [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"}) by multiplying and normalizing: \[P(t|k,A) \propto \mathcal{L}(t|k,A)P(t)\] Our estimates are consistent with the method of in \(\sim 90\%\) of cases (Figure [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}). In every case, our strong prior forces our central estimate to be older than the central estimate from the method (Figure [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}). Almost all AHi subunits are consistent at the \(95\%\) confidence level with being Hesperian or Amazonian in age, however (considering only crater data and setting aside cross-cutting relationships) most of the AHi subunits are also consistent with being Noachian in age. Using the posterior age CDF's for each AHi subunit, we derived the CDF for the age of the youngest fan-bearing AHi subunit. To do so, we assumed the ages of each AHi subunit are independent random variables: \[P(t_{min} \geq t_{ref}|t_{ref}) = \prod_i P(t_i \geq t_{ref}|k_i,A_i, t_{ref})\] where \(i\) is an arbitrary index over the fan-bearing AHi subunits. We computed this distribution for two populations: (a) all fan-bearing AHi subunits (N = 70), and (b) those fan-bearing AHi subunits where localized impact trigger is excluded as an explanation for alluvial fan formation (N = 13). Population (b) comprises those fan-bearing subunits in which pre-or syn-fluvial impact craters were identified. The results are summarized in Figure [\[fig:fig4\]](#fig:fig4){reference-type="ref" reference="fig:fig4"}. We found at the \(95\%\) confidence level that \(t_{min}\lesssim 0.9\) Ga if we include all fan-bearing AHi subunits. Considering only fan-bearing AHi subunits where localized impact trigger is excluded, we found at the \(95\%\) confidence level that \(t_{min}\lesssim 2.5\) Ga.
## Modeling Fan Formation Scenarios
To investigate the formation timing of alluvial fans in a way that incorporates information from crater-counts on non-fan-bearing AHi subunits, we developed two simple statistical models (Figure [\[fig:fig5\]](#fig:fig5){reference-type="ref" reference="fig:fig5"}). In the simplest model (pulse-formation model), we assumed that alluvial fans formed at a single moment in history, \(\tau_{form}\), and that all AHi subunits with age \(t_i \geq \tau_{form}\), had an equal probability, \(p_{max}\), of acquiring alluvial fans. In the second model (prolonged-formation model), each AHi subunit had a constant instantaneous probability, \(p_{max}/\Delta\tau\), of acquiring alluvial fans during the \(\Delta\tau\)-long time period ending at \(\tau_{end}\). This prolonged-formation model is approximately equivalent to a model with many short bursts of activity that are evenly spaced or randomly spaced throughout the \(\Delta\tau\)-long time period ending at \(\tau_{end}\). In both models, the probability of an AHi subunit younger than \(\tau_{end}\) or \(\tau_{form}\) acquiring fans was assumed to be 0. Both models are motivated by searching for changes in the probability of acquiring fans as a function of time, for example due to atmospheric evolution. Neither model is set up to identify preferred spatial patches for fan formation, for example latitude dependent climate variations. Any such variations are averaged out in the globally uniform (but time-varying) probability of acquiring alluvial fans. For both statistical models, the formation scenarios can be viewed both in an instantaneous framework (left panel, Figure [\[fig:fig5\]](#fig:fig5){reference-type="ref" reference="fig:fig5"}), and in a cumulative framework (right panel, Figure [\[fig:fig5\]](#fig:fig5){reference-type="ref" reference="fig:fig5"}). Both viewpoints are instructive, as the instantaneous view relates to a hypothetical climate scenario (e.g. short-lived, global transient greenhouse), while the cumulative view relates fan-acquisition probabilities to AHi subunit ages, \(t_i\). The relationship between these two views can be formalized mathematically via simple integrals for both models: \[\mathcal{L}(\tau_{form},p_{max}|F_i = 1, t_i) = \int_0^{t_i} P(F \rightarrow 1 | \tau_{form}, p_{max}, \tau)d\tau\] \[\mathcal{L}(\tau_{end},\Delta\tau,p_{max}|F_i = 1, t_i) = \int_0^{t_i} P(F \rightarrow 1 | \tau_{end},\Delta\tau, p_{max}, \tau)d\tau\] where \(F_i=1\) when the \(i^{th}\) AHi subunit contains fans (0 otherwise), and \(P(F \rightarrow 1)\) is the probability of acquiring fans, which varies with time (Figure [\[fig:fig5\]](#fig:fig5){reference-type="ref" reference="fig:fig5"}). Note, because \(F_i\) is a binary variable, \(P(F_i=0)=1-P(F_i=1)\), which defines \(\mathcal{L}\) in the case that \(F_i=0\). Because we only know the AHi subunit ages, \(t_i\), probabilistically, we computed expected likelihoods (instead of the likelihoods themselves) as a function of model parameters and crater density data by marginalizing over the age posterior distributions: \[E(\mathcal{L}(\tau_{form},p_{max}|F_i, k_i,A_i)) = \int_0^{4.5}\mathcal{L}(\tau_{form},p_{max}|F_i, t)P(t|k_i,A_i)dt\] \[E(\mathcal{L}(\tau_{end},\Delta\tau,p_{max}|F_i, k_i,A_i)) = \int_0^{4.5}\mathcal{L}(\tau_{end},\Delta\tau,p_{max}|F_i, t)P(t|k_i,A_i)dt\] where \(k_i\) is the observed crater count and \(A_i\) the count region area for the i'th (out of 212) AHi subunits. Taking the data from every AHi subunit together, we can multiply the expected likelihoods together to estimate the overall expected likelihood as a function of model parameters and data: \[E(\mathcal{L}(\tau_{form},p_{max}|\vec{F}, \vec{k},\vec{A})) = \prod_i E(\mathcal{L}(\tau_{form},p_{max}|F_i, k_i,A_i))\] \[E(\mathcal{L}(\tau_{end},\Delta\tau,p_{max}|\vec{F}, \vec{k},\vec{A})) = \prod_i E(\mathcal{L}(\tau_{end},\Delta\tau,p_{max}|F_i, k_i,A_i))\] We computed the expected likelihood for the pulse-formation model as a function of \(\tau_{form}\) and \(p_{max}\), allowing them to vary in \[0 Ga, 4 Ga\] and \[0,1\], respectively (Figure [\[fig:fig6\]](#fig:fig6){reference-type="ref" reference="fig:fig6"}). The model achieved maximum likelihood at \(\tau_{form}=1.5\) Ga and \(p_{max} = 0.36\). Assuming a uniform prior over \(p_{max}\), we marginalized the likelihoods over \(p_{max}\) ("summed over \(p_{max}\)") to estimate a posterior PDF for \(\tau_{form}\) (right panel of Figure [\[fig:fig6\]](#fig:fig6){reference-type="ref" reference="fig:fig6"}). The \(95\%\) confidence level for the estimated posterior PDF in the right panel of Figure [\[fig:fig6\]](#fig:fig6){reference-type="ref" reference="fig:fig6"} is slightly more restrictive than the right-most extent of the middle red contour in the right panel of Figure [\[fig:fig6\]](#fig:fig6){reference-type="ref" reference="fig:fig6"}-this is due to the marginalization over \(p_{max}\). At the \(95\%\) confidence level, we found that in the pulse-formation scenario, the data requires fan formation to occur within the last \(\sim 2.8\) Gyr. We computed the expected likelihood for the prolonged-formation model, as a function of \(\tau_{end}\), \(\Delta\tau\), and \(p_{max}\), allowing them to vary in \[0 Ga, 4 Ga\], \[0, 4 Ga \(- \tau_{end}\)\], and \[0,1\], respectively (Figure [\[fig:fig7\]](#fig:fig7){reference-type="ref" reference="fig:fig7"}). The model achieves maximum likelihood at \(\tau_{end}= 0\) Ga, \(\Delta\tau=3\) Ga, \(p_{max}=0.36\) (Figure [\[fig:fig7\]](#fig:fig7){reference-type="ref" reference="fig:fig7"}, top left panel). We compared the maximum likelihoods from the two models (this comparison does not require assuming that either model is a good model). Despite a maximum likelihood solution with \(\Delta\tau>0\), the difference in maximum likelihoods between the prolonged-formation model and the pulse-formation model (equivalent to the prolonged-formation model when \(\Delta\tau=0\)) was not sufficiently large to reject the pulse-formation model via a likelihood-ratio test (\(p < 0.81\), where our computed test statistic, \(\lambda_{LR} \approx 0.06\), is given by twice the difference in the maximum log-likelihoods for each model and is asymptotically chi-square distributed with 1 degree of freedom). Assuming uniform priors, we marginalized the likelihoods over \(\Delta\tau\) and \(p_{max}\) to estimate a posterior distribution for \(\tau_{end}\) (Figure [\[fig:fig7\]](#fig:fig7){reference-type="ref" reference="fig:fig7"}). At the \(95\%\) confidence level, we found that in the prolonged-formation scenario, our dataset requires fan formation to have persisted into the last \(\sim 2.3\) Gyr.
# Discussion
Using the crater counts on AHi subunits' ejecta blankets, as well as some simple statistical models, we were able to place constraints on the timing of the formation of alluvial fans. Our estimate of the CDF for the age of the youngest fan indicated that at least one fan likely formed by a process other than localized impact trigger in the last \(\sim 2.5\) Gyr. Further, our CDF estimate indicates that at least one fan likely formed in the last \(\sim 0.9\) Gyr. This is broadly consistent with the results from the pulse-formation model (which incorporates information from crater-counts on non-fan-bearing AHi subunits), in which the best-fitting time of fan formation was \(\sim 1.5\) Gyr. This indicates that fan formation likely persisted well into the Amazonian epoch, consistent with and. Although the pulse-formation model does not favor formation occurring in the last \(\sim 1\) Gyr (due to the fact that very young values of \(\tau_{form}\) force non-fan-bearing AHi subunits to take on likelihood \(1-p_{max}\) instead of 1 with too high a probability), it does not explicitly reject it. Further, at least some of the young fans can be explained by localized-impact trigger, which is not permitted in the pulse-formation model scenario. We estimated parameters for a more complex model, the prolonged-formation model, and compared the results with the pulse-formation model. The introduction of a third parameter introduced a large degeneracy between \(\tau_{end}\) and \(\Delta\tau\) in our parameter retrieval (Figure [\[fig:fig7\]](#fig:fig7){reference-type="ref" reference="fig:fig7"}). Although the model selected \(\tau_{end}=0\) Ga with fan-formation lasting several Gyr, a likelihood ratio test revealed that the maximum likelihood is not sufficiently higher in the prolonged-formation model than in the pulse-formation model to definitively justify the addition of the third parameter. Thus we cannot reject either the pulse-formation or prolonged-formation model. Further, we do not have tight constraints on \(\tau_{end}+\Delta\tau\), due to the degeneracy in our prolonged-formation model parameter retrieval. As a result, we cannot conclusively say, using this crater data alone, whether the alluvial fans formed in an era entirely separate from the valley networks (i.e. \(\tau_{end}+\Delta\tau \lesssim 3.4\) Ga with high confidence), or if the alluvial fans represent a long tail of continuous fluvial activity (i.e. \(\tau_{end}+\Delta\tau \gtrsim 3.4\) Ga with high confidence). For both models, we showed marginalized distributions for \(\tau_{form}\) and \(\tau_{end}\), assuming uniform priors over the other parameters, \(p_{max}\) and \(\Delta\tau\). Unlike the prior distributions for the ages of the AHi subunits, we had no particular reason to choose these parameter prior distributions (although we note that marginalizing over \(p_{max}\) is likely okay due to its strong determination by the data, see Figure [\[fig:fig7\]](#fig:fig7){reference-type="ref" reference="fig:fig7"}). The marginalized distributions should thus be viewed only as condensed data visualizations, rather than true posterior distributions. Despite this, maximum likelihood estimates from both models indicate fan formation continued into the last \(\sim 1.5\) Gyr. This global age is substantially (\(\sim 2\) Gyr) younger than existing quantitative estimates of the age of fan formation from a study of two regions. This can be explained by the fact that our model allows for spatially patchy fan formation (i.e. \(p_{max} < 1\)) while the method of assumes uniform fan formation and thus would overestimate the young limit on alluvial fan formation if applied to our dataset. We note that the assumption that \(p_{max}=1\) may be perfectly acceptable within the study areas. Our results are consistent with young age estimates obtained using smaller count areas for some fans (e.g. @GRANT2011 [@GRANT2019]), and are also consistent with Amazonian age estimates made for some non-fan features that indicate flowing surface liquid water (e.g., @DICKSON2009 [@FASSETT2010; @HOWARDMOORE2011; @WEITZ2013; @EHLMANNBUZ2015; @BROSSIER2021; @BUTCHER2021]). A K-Ar age of (2.12 \(\pm\) 0.36) Ga obtained for aqueous minerals at Gale crater by the *Curiosity* rover is consistent with our scenario but does not require it. Our strategy in this study does not incorporate the prior indications of young activity from these previous studies, and our crater data does not overlap with that used by these previous studies, therefore our results are independent. By themselves, our data do not allow us to determine whether fan formation was intermittent (either synchronous between fan-forming regions, or regionally diachronous), or alternatively if water flow occurred in more-or-less every year in all regions where fans are present. Consensus on this important topic may have to await a future landed mission. In general, data availability limited our parameter retrievals for \(\tau_{form}\) and \(\tau_{end}\), which had large uncertainties associated with them. Indeed, our crater counts were sparse due to our lower diameter cutoff of 4 km, causing our AHi subunit age posteriors to be largely influenced by our choice of prior distribution and have high variance (Figures [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} and [\[fig:fig3\]](#fig:fig3){reference-type="ref" reference="fig:fig3"}). We performed a series of sensitivity tests to illustrate how the model performs both with more data and with idealized data. First, we performed the parameter estimation with the same data, but with the outer annulus cutoff removed (Figure [\[fig:fig8\]](#fig:fig8){reference-type="ref" reference="fig:fig8"}). This had the effect of biasing our retrievals to older ages due to the larger influence of old craters "poking-through" the ejecta blankets near the edges (Figure [\[fig:fig8\]](#fig:fig8){reference-type="ref" reference="fig:fig8"}). Indeed, the pulse-formation model achieved maximum likelihood at \(\tau_{form} = 1.7\) Ga and \(p_{max} = 0.36\). Further, the increased quantity of data decreased the variance of our estimates, which is evident in the decreased degeneracy in the prolonged-formation model (Figure [\[fig:fig8\]](#fig:fig8){reference-type="ref" reference="fig:fig8"}). Next, in order to evaluate how effectively the mathematical model can retrieve parameter values, we carried out a test using synthetic data. Specifically, we performed the parameter retrievals using the original count areas (including outer annulus limit), using synthetic data where the absence or presence of alluvial fans and ejecta crater counts were randomly generated assuming that the pulse-formation model was true, that the fans formed at 1.5 Ga with probability 0.5, and that AHi subunits have ages drawn from the prior distribution, truncated at 4 Ga. In general, the model performs well and achieves maximum likelihood within \(1\sigma\) of the correct value (Figure [\[fig:fig9\]](#fig:fig9){reference-type="ref" reference="fig:fig9"}). However, there is still substantial variance in the results, and the model is unable to substantially break the degeneracy between the end of fan formation and the duration of the fan formation era, although it does achieve maximum likelihood at \(\tau_{end} = 2.1\) Ga, \(p_{max} = 0.47\) and \(\Delta\tau = 100\) Myr (Figure [\[fig:fig9\]](#fig:fig9){reference-type="ref" reference="fig:fig9"}). This suggests that more craters are needed to constrain the formation timing. We repeated the experiment with synthetic data, artificially increasing the AHi count areas by a factor of 10. This had the effect of placing roughly 10 times as many craters on each subunit, thereby decreasing the uncertainty in the posterior age distributions. As a result, the variance in our parameter retrievals shrunk (compare Figure [\[fig:fig10\]](#fig:fig10){reference-type="ref" reference="fig:fig10"} with Figure [\[fig:fig9\]](#fig:fig9){reference-type="ref" reference="fig:fig9"}). Further, the parameter retrieval has a slightly increased ability to discriminate between the prolonged-and pulse-formation models (relative to the synthetic data experiment with uninflated AHi subunit areas) and preferentially selects a fan formation era duration of 0 (Figure [\[fig:fig10\]](#fig:fig10){reference-type="ref" reference="fig:fig10"}). This indicates that in order to obtain better estimates for the timing and duration of alluvial fan formation, one needs substantially more data and must consider craters at smaller diameters (recall our lower cutoff was 4 km). Although we ignored the possibility that surface processes efficiently obliterated 4 km diameter craters in the post-Noachian, one must correct for crater obliteration to avoid underestimating age when smaller craters are considered. In our methodology, we did not account for uncertainty in the chronology function, which is continually being re-evaluated from a combination of crater counting, geochronology, and solar system dynamics perspectives. It is possible that our age estimates are systematically younger or older than the true ages, which would have biased our fan formation timing estimates. This potential effect is especially important in the data-limited case, where the prior strongly informs posterior age distributions (e.g. if there were ample data, the posterior curves in Figure [\[fig:fig2\]](#fig:fig2){reference-type="ref" reference="fig:fig2"} would coincide with the normalized likelihood curves). A related point is that if we had imposed an exponentially decaying prior likelihood on fan formation time, then our fan age estimates would have been older. Further, we assumed a maximum age of 4.5 Ga in our prior age distribution. In principle, one could impose a younger maximum age based on stratigraphy to obtain tighter constraints. We re-performed our parameter retrieval assuming a maximum age of 3.54 Ga (approximately the Noachian/Hesperian boundary). However, roughly half of the AHi subunits achieved maximum posterior probability at 3.54 Ga, indicating that this prior was too restrictive. It is possible that with improved crater surveys, the data would produce maximum a-posteriori ages that are well into the Amazonian. Finally, we did not account for the possibility that the fan catalog is incomplete. It is possible that some fans were wrongly identified and that other fans were missed in certain craters in the survey. To explore this effect, we performed a sensitivity test where the probability of an AHi subunit acquiring fans after the end of fan formation (or after the pulse of fan formation) was 0.001, instead of 0. This had a negligible effect on the outcome, and we concluded that survey incompleteness needs to be more severe (e.g. a \(10\%\) chance of missing all the fans in a crater) to substantially alter the results.
# Conclusions
We combined a global geologic map, a global database of Martian impact craters, a global database of alluvial fans, and statistical models to make inferences about the timing and duration of alluvial fan formation. Although differing in detail due to differing assumptions, the results of all our models suggest that fan formation had to have persisted into the last \(\sim 2.5\) Gyr (based on our most robust and conservative estimate, see Figure [\[fig:fig4\]](#fig:fig4){reference-type="ref" reference="fig:fig4"}), well into the Amazonian time period. This estimate is \(\sim 1\) Gyr younger than previous estimates. Overall, our modeling procedure was not able to constrain the duration of fan formation. As a result, we were unable to determine from this dataset whether the alluvial fans formed in a long tail of continuous fluvial activity that overlapped with valley-network formation, or if they formed in an era entirely separate from the valley networks. In general, our inferences were limited by a lack of crater count data, and our tests with synthetic data suggest that detailed surveys of smaller craters on AHi subunit ejecta blankets will enable much tighter constraints on fan formation. Going forward, we hypothesize that more data and more complex statistical models (i.e. ones with more parameters) for the population of alluvial fans (or other geomorphic features) will enable determination of both (a) their joint temporal and spatial evolution and (b) the relative importance of climate-triggered and localized impact-triggered fan formation. The method we have developed can also be applied to determining the ages of other crater-hosted features on Mars. Unraveling the evolution of alluvial fan (and/or other crater-hosted fluvial features') formation would thereby provide even stronger constraints on the evolution of Mars' climate. We thank Stuart Robbins and Marisa Palucis for insightful discussions that helped conceive this study. We thank two reviewers for useful comments. We acknowledge funding from NASA grants 80NSSC20K0144, NNX16AJ38G, and NNX15AM49G. | {'timestamp': '2021-09-14T02:18:35', 'yymm': '2109', 'arxiv_id': '2109.05407', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05407'} |
# Introduction
The current pandemic revitalized research on group testing, a methodology to reduce the number of required tests to screen a large population for a certain disease. The increased testing efficiency results from jointly testing groups, instead of subjecting each individual to a test. Specifically, we consider the scenario of *probabilistic* group testing as first described by, where every individual has a certain (known) probability to be infected. Probabilistic group testing is in contrast to combinatorial group testing where one assumes a fixed known number of infected individuals. Furthermore, we assume that while the number of tests is limited, all tests are perfectly accurate. Although the case of imprecise tests also leads to interesting resource allocation problems, we will not consider it in this work. Thus, if none of the individuals in a tested group are infected, the test will yield a negative result with certainty and one test was sufficient to obtain a definite result for several individuals. If, however, the test is positive, one cannot say which individuals in the group are infected and further tests have to be conducted. Formally, a *testing strategy* is a deterministic algorithm, describing which groups of individuals are pooled and jointly subjected to the test. We assume that the choice of the next group can depend on the results of previously conducted tests. This *adaptive* group testing, is in contrast to non-adaptive group testing (for a recent survey see ), where the testing strategy is fixed and all tests can thus be performed in parallel.
## Previous work {#previous-work .unnumbered}
The theoretical work on adaptive, probabilistic group testing started with who considered the simple strategy of testing groups of \(M\) individuals and subsequently, if the pooled test is positive, testing each individual in the group separately. Although this strategy is far from optimal, it is easy to implement and can significantly increase the testing capabilities, compared to testing every sample individually. Further research in this direction was focused on finding new, more efficient and practical testing strategies, e.g., nested testing, binary splitting, and array testing. The binary splitting algorithms by perform close to the theoretic optimum, within a factor of \(1.11\) of an information-theoretic lower bound, as shown by. However, such a binary splitting technique requires extensive bookkeeping and many tests need to be performed sequentially, and cannot be conducted in parallel. Simpler, two-stage procedures are investigated by. The close relation of adaptive, probabilistic group testing to variable length source coding is well-detailed by . Due to the current pandemic, several works rediscover slight variations of these results and argue for the use of group testing. Also, the practical implementation of group testing has seen a new surge of research, in particular, questioning the practical use in testing for SARS-CoV-2. Here, however, the focus was on the classical Dorfman testing strategy and only few works considered more elaborate non-adaptive testing strategies. Hardly any works go beyond the assumptions of perfect sensitivity and independence of the infection status of tested individuals. However, takes the dilution effect into account, i.e., reduced sensitivity of tests for large groups, and performed a simulation study that found benefits of group testing if there is positive correlation of infection status between individuals in the same group. The common ground of all works above is that they focus on identifying exactly, which individuals are infected. This is a valid strategy and clearly the best outcome. However, there can be situations where not sufficiently many tests are available to subject all potentially infected individuals to a test. This is particularly the case when many individuals are (potentially) infected and testing resources are scarce. Here, a strategy needs to be found that uses these limited resources effectively. When insufficient tests for the entire population are available, it is unavoidable that some healthy individuals will be treated as infected and/or some infected individuals treated as healthy. For the design of a testing strategy, it is important to note, that these two events are not equally harmful in general. It might be considerably worse to have an undetected infection (false negative) present, than to treat one healthy individual as infected (false positive). A suitable balance has to be found and sensitive questions like "How many false positives are we willing to accept to prevent one false negative?" need to be answered to do so. Here, assuming that quantitative answers to these questions can be given, we propose a framework for designing and evaluating group testing strategies. Mathematically, the resulting problem is one of rate-distortion theory, a branch of information theory, established in the 1950s in a seminal paper by . This connection allows us to formulate bounds on the performance of any group testing strategy. Although the problem of (combinatorial) group testing was extensively explored by information-theorists (e.g., in Ch. 24--29 in and by ), this rate-distortion viewpoint has apparently not been considered so far. Thus, fundamental bounds are missing even for elementary scenarios. Only recently, researchers suggested that "the scarcity of tests obviously means that it is better to use a test to detect the virus in another untested group than to try to discover who is infected in a positive group" which is a first step into the direction of the rigorous theory developed here. Finally, the scenario discussed in is closest to our ideas but considers only a single fixed testing strategy. We will use it in our theory for comparison and as a starting point for some more evolved testing strategies.
# Problem formulation {#sec_problem}
We assume that we have a sequence of individuals and the infection status of the \(n\)-th individual is given by a binary random variable \(X_n\) on \(\Omega := \{0,1\}\), where \(X_n=1\) corresponds to being infected and \(X_n=0\) to being healthy. All \(X_n\) are assumed to be independent but not necessarily identically distributed. Thus, each \(X_n\) is a Bernoulli(\(p_n\)) random variable with possibly different probability \(\Pr[X_n=1] = p_n \in (0, 1)\). The second ingredient we need is a *cost function* \(\rho_n(x_n, y_n)\) that models the cost of wrong assignments, i.e., assigning an estimated infection status \(y_n\) to the \(n\)-th individual with actual infection status \(x_n\). In contrast to communication scenarios where \(0\) and \(1\) are usually interchangeable, the cost \(\rho_n(0,1)\) of false positive assignment (i.e., a healthy individual is wrongly assigned an infected status) is not necessarily the same as the cost \(\rho_n(1,0)\) of a false negative assignment (i.e., an infected individual is wrongly assigned a healthy status). Thus, we define \[\begin{aligned}
\rho_n(0,1) = b_n\,, && \text{ and} && \rho_n(1,0) = c_n\,, \label{eq:false} \end{aligned}\] where \(b_n, c_n > 0\). The cost of correct assignments is set to zero, i.e., \(\rho_n(0,0) = \rho_n(1,1) = 0\) and the total cost \(\rho\colon \Omega^N \times \Omega^N \to \mathbb{R}\) is given by summation \((\mathbf{x},\mathbf{y}) \mapsto \rho(\mathbf{x}, \mathbf{y}) = \sum_{n=1}^{N} \rho_n(x_n, y_n)\). We now turn to the mathematical description of testing strategies. A *testing strategy* for \(N\) individuals consists of a test procedure and a decision procedure. An \((N,K)\)-*test procedure* is given by \(\boldsymbol{\eta} = (\eta_1, \eta_2, \dots, \eta_K)\) where \(\eta_1 \in \mathcal{P}\big(\{1,2,\dots,N\}\big)\) and for \(k > 1\), \(\eta_k\colon \Omega^{k-1} \to \mathcal{P}\big(\{1,2,\dots,N\}\big)\), where \(\mathcal{P}\big(\{1,2,\dots,N\}\big)\) denotes the collection of all subsets of \(\{1,2,\dots,N\}\). Here, the set \(\eta_1\) indicates the group used for the first test and the set-valued function \(\eta_k\) indicates the group used for the \(k\)-th test, given the results of the previous \(k-1\) tests. The corresponding *test function* \(\boldsymbol{\vartheta}\colon \Omega^N \to \Omega^K\) is given by \(\vartheta_1(\mathbf{x}) = \max\{x_n | n \in \eta_1\}\) and for \(k > 1\), we have \(\vartheta_k(\mathbf{x}) = \max\big\{x_n \big| n \in \eta_{k}\big(\vartheta_1(\mathbf{x}), \vartheta_2(\mathbf{x}), \dots, \vartheta_{k-1}(\mathbf{x})\big)\big\}\). Thus, the \(k\)-th component \(\vartheta_k\) corresponds to the result of the \(k\)-th test. A \((K,N)\)-*decision procedure* is a mapping \(\kappa\colon \Omega^K \to \Omega^N\) that assigns, based on the outcome of \(K\) tests, a status (infected/healthy) to all \(N\) individuals. The concatenation of \(\boldsymbol{\vartheta}\) and \(\kappa\) maps the true status \(\mathbf{X} := (X_1, \dots, X_N)\) of all \(N\) individuals to an estimated status \((Y_1, \dots, Y_N) = \kappa \big(\boldsymbol{\vartheta} (X_1, \dots, X_N) \big)\) of these individuals using \(K\) group tests. In total, this corresponds to \(R=K/N\) [TpI]{acronym-label="TpI" acronym-form="singular+short"} which is referred to as the *rate* of the testing strategy. If \(R < 1\), which is the regime we are interested in, there is a positive probability that the tests will not enable a perfect identification of all infected individuals. Note that such an estimate (\(0\): not infected, \(1\): infected) has to be given for all \(N\) individuals. We do not allow for individuals to be "skipped," which would correspond to a ternary output alphabet. To assess the average cost of the wrong assignments for given test and decision procedures, we use the *expected cost per individual*, defined as the expected value \[D_{\text{test}}^{(\mathbf{X}, \rho)}(\boldsymbol{\vartheta}, \kappa) = \mathbb{E}\bigg[\frac{1}{N}\sum_{n=1}^N \rho_n(X_n, Y_n)\bigg].\] Because there are only finitely many possible choices for \(\boldsymbol{\vartheta}\) and \(\kappa\), we can define the minimum \[\label{eq:defdfink} D^{(\mathbf{X}, \rho)}(K) = \min_{\boldsymbol{\vartheta}, \kappa} D_{\text{test}}^{(\mathbf{X}, \rho)}(\boldsymbol{\vartheta}, \kappa),\] where \(\boldsymbol{\vartheta}\) and \(\kappa\) range over all \((N,K)\)-test and \((K,N)\)-decision procedures, respectively. The quantity \(D^{(\mathbf{X}, \rho)}(K)\) specifies the minimal cost, measured by \(\rho\), that can be achieved by using \(K\) tests for the \(N\) individuals with random infection status \(\mathbf X\).
# Results {#sec:results}
Calculating \(D^{(\mathbf{X}, \rho)}(K)\) directly from [\[eq:defdfink\]](#eq:defdfink){reference-type="ref" reference="eq:defdfink"} is computationally infeasible, unless \(N\) and \(K\) are very small. Nevertheless, we can use information-theoretic ideas to provide bounds. These bounds are based on the idea of keeping the rate \(R = K/N\) fixed, while letting \(N\) approach infinity. Our first result is a lower bound, that holds if all individuals share the same parameters. Here, and in the remainder of the paper, we use the symbol \(H_2(p) =-p\log p-(1-p)\log(1-p)\) for the binary entropy function, \(\log(\,\cdot\,)\) denotes the logarithm to base \(2\), and we adopt the convention that "\(0\cdot\log 0 = 0\)." The proofs of the results in this are presented in [5](#sec:itdetails){reference-type="ref" reference="sec:itdetails"} and in the appendices. The following presents a lower bound on \(D^{(\mathbf{X}, \rho)}(K)\) for the case when infection is equally likely and independent across the entire population. It follows immediately from the more general [\[th:partitions\]](#th:partitions){reference-type="ref" reference="th:partitions"}, which will be stated later in this .
We emphasize that, in contrast to similar results in classical information theory, the lower bound in [\[thm:RD_function\]](#thm:RD_function){reference-type="ref" reference="thm:RD_function"} cannot always be achieved arbitrarily closely for increasing \(N\). For example, for \(p = \frac{1}{2}(3-\sqrt{5}) \approx 0.381\) and \(v=0\), we obtain \(\bar D(p, a, v)=0\) and \(\bar R(p, a, v)=H_2(p)\approx 0.959\), although it is known that only individual testing (i.e., \(R=1\)) can achieve \(D^{(\mathbf{X}, \rho)}(K)=0\) in this setting. Even though the lower bound in [\[thm:RD_function\]](#thm:RD_function){reference-type="ref" reference="thm:RD_function"} is somewhat cumbersome and difficult to grasp intuitively, it can easily be calculated for a given scenario. We will compare it to proposed testing strategies in [\[fig:rda50\]](#fig:rda50){reference-type="ref" reference="fig:rda50"} to illustrate its applicability. We next calculate the necessary number of tests and the resulting cost for some simple testing strategies. We first consider the strategy, proposed by, to separate the population into equally sized and disjoint groups, test each group, and assign an infected status to each member of a positive group without conducting further tests. We refer to this strategy as *one stage group testing (1SG)* and designate the size of the group in parenthesis, e.g., 1SG(\(50\)) for a group size of \(50\) individuals. We restate the following simple result from Section 2.3.1 in, which is a special case of the more general [\[lem:ksg\]](#lem:ksg){reference-type="ref" reference="lem:ksg"} below.
Note that the exact rate and expected cost can be achieved only for a population \(N\) that is an integer multiple of the group size \(u\). However, the overhead is at most \(1\) additional test for a final group of smaller size, resulting in an additional \(1/N\) [TpI]{acronym-label="TpI" acronym-form="singular+short"} and a negligible decrease of the expected cost. Since we are usually concerned with large \(N\), we ignore these terms in this and the following testing strategies. Of course, 1SG(\(u_1\)) can be further extended by testing those individuals again, that belong to a positively tested group of size \(u_1\). Specifically, we can separate the group into disjoint subgroups of smaller size \(u_2\) and subject these subgroups again to a group test. This *two stage group testing (2SG)* strategy is a generalization of Dorfman testing which corresponds to the case \(u_2 = 1\), i.e., each individual in a positive group is tested separately. The decision strategy remains the same as in 1SG: Those individuals that belong to a positively tested subgroup (of size \(u_2\)) are declared infected. We again designate the group sizes in parenthesis, e.g., 2SG(\(50,10\)) for a group size of \(50\) individuals that is divided into five subgroups of \(10\) individuals each, if the first group test is positive. Again, we can calculate the rate and expected cost in closed form.
Evidently, more than 2 stages could be used, but even more bookkeeping is then required and, typically, very large group sizes are needed to obtain a benefit. Thus, this extension may not be practically useful anymore. For the sake of completeness, we nevertheless provide the rate and expected cost of \(k\) stage group testing, abbreviated as \(k\)SG(\(u_1, \dots, u_k\)), where \(u_\ell\) denotes the group size at stage \(\ell\). Here, positive groups of size \(u_\ell\) are separated into smaller subgroups of size \(u_{\ell+1}\) for \(\ell=1, \dots, k-1\) and individuals that belong to a positive subgroup at stage \(k\) are declared infected, while all others are declared healthy. A proof of the result is given in [\[app:proofksg\]](#app:proofksg){reference-type="ref" reference="app:proofksg"}.
The example above illustrates that the proposed strategies, although very simple, can outperform the best known testing strategies if there are not sufficient tests available to test all individuals. This is because previous strategies always aimed at exactly identifying the infection status and did not consider the possibility to declare an infection status based on imperfect information. However, there are also many situations where the simple strategies we discussed above are useless. In particular, if the relative cost \(a\) is small, then it is hardly ever useful to declare an individual infected if one is not very sure about the infection status. We now turn to the more general setting of a heterogeneous population, i.e., individuals may have different prevalence \(p_n\) and also different costs \(b_n\) and \(c_n\). One simple example of this situation is to distinguish between individuals with symptoms (that have a higher prevalence \(p_n\)) and individuals without symptoms. Additionally, different costs may occur for individuals in critical areas, e.g., health care. There, a higher risk of infecting vulnerable individuals may increase the cost \(c_n\) of false negatives, and additionally, a higher false positive cost \(b_n\) might be incurred due to the importance of the work being performed by these individuals. We will discuss a specific scenario in [4](#sec:case){reference-type="ref" reference="sec:case"} but first present the lower bound, an extension of [\[thm:RD_function\]](#thm:RD_function){reference-type="ref" reference="thm:RD_function"}. The proof of the following will be presented at the end of [5](#sec:itdetails){reference-type="ref" reference="sec:itdetails"}, as it requires several results from rate-distortion theory which are introduced there.
# Historic example case {#sec:case}
We consider the SARS-CoV2 pandemic situation in Austria in mid November 2020. On average there were \(N=8\,916\,845\) individuals living in Austria in 2020 . During the three days from 12th of November 2020 until 14th of November 2020 a total of \(N_t = 103\,621\) individuals were subjected to a PCR-test for SARS-CoV2 . Of these \(N_t\) individuals, \(20\,349\) tested positive . This corresponds to a prevalence of \(p_t = 0.196\) in this tested subpopulation. We will refer to this subpopulation as the *high prevalence subpopulation*. At the same time, a prevalence study found that approximately \(3.1\%\) of the total population were infected . Thus, in the untested population of \(N_u = 8\,813\,224\) the prevalence was about \(p_u = 0.029\). We will refer to this subpopulation as the *low prevalence subpopulation*. To illustrate the full potential of our model, we further consider individuals working in health care separately and will assign a higher cost for wrong assignments within this subpopulation. The most recent count of health care professionals working in hospitals and health care centers in Austria was \(N_h = 121\,567\) at the end of 2019 . Note that a finer separation into subpopulations is of course possible but avoided for the sake of simplicity. It is difficult to argue for the choice of specific costs of wrong assignments. However, in mid November 2020, a lockdown was issued in Austria,[^1] which we interpret as the turning point where considering all (untested) individuals as infected is less costly than considering these individuals as healthy. This implies \((1-p) b \approx p c\), thus, at a prevalence of \(p = 0.029\) in the untested, general population, we obtain \(c \approx 33 b\). We normalize \(b=1\) and choose \(c = 33\). Since health-care facilities were not closed during this time, we assume that individuals working in health care have a different trade-off between \(b\) and \(c\). Specifically, we assume that a false positive assignment for this subpopulation is significantly more expensive and we set \(b=6\) for these individuals. Although also a different \(c\) value could be argued, we assume that the professional training counters the higher risk due to closer contact with susceptible individuals and we keep \(c\) the same. Since there is no data available on the prevalence within the health care system, we assume that it is the same as in the general population. In particular, also the split into the high prevalence and low prevalence subpopulations is assumed to be the same. We thus end up with the following four subpopulations
- The first subpopulation consists of individuals working in health care that belong to the high prevalence subpopulation with \(p^{(1)}=0.196\) and costs \(b^{(1)}=6\), \(c^{(1)} = 33\). We assume that \(N^{(1)} = 1\,413\) belong to this subpopulation.
- The second subpopulation consists of individuals working in health care that belong to the low prevalence subpopulation with \(p^{(2)}=0.029\) and costs \(b^{(2)}=6\), \(c^{(2)} = 33\). We assume that \(N^{(2)} = 120\,154\) belong to this subpopulation.
- The third subpopulation consists of individuals not working in health care that belong to the high prevalence subpopulation with \(p^{(3)}=0.196\) and costs \(b^{(3)}=1\), \(c^{(3)} = 33\). We assume that \(N^{(3)} = 102\,208\) belong to this subpopulation.
- The fourth subpopulation consists of individuals not working in health care that belong to the low prevalence subpopulation with \(p^{(4)}=0.029\) and costs \(b^{(4)}=1\), \(c^{(4)} = 33\). We assume that \(N^{(4)} = 8\,693\,070\) belong to this subpopulation.
Our lower bound and the performance of the optimal combinations of 1SG and 2SG strategies for these specific parameters are illustrated in [\[fig:subpop\]](#fig:subpop){reference-type="ref" reference="fig:subpop"} in comparison to optimized binary splitting algorithms and individual testing. More specifically, the depicted curves correspond to the optimal use of the available tests under the given strategy. For individual testing and the binary splitting algorithms, this means that depending on the number of tests, a certain amount of people can be tested and correctly informed about their status, while many untested individuals remain and are assigned a status without being tested resulting in the depicted cost. For the \(103\,621\) tests used during the discussed period, which correspond to \(0.0116\) [TpI]{acronym-label="TpI" acronym-form="singular+short"}, the optimized strategy is to subject as many individuals from the fourth subpopulation as possible to a test using a 1SG(\(33\)) testing strategy. By this approach, the expected cost is \(0.816\). Individual testing could only reduce the expected cost to \(0.944\), hardly any improvement over not testing at all at an expected cost of \(0.956\). Note that in all subpopulations and strategies we assume that an optimal decision for untested individuals is used, namely, that a healthy status is assigned to untested individuals in subpopulations two and four and an infected status to untested individuals in subpopulations one and three. To illustrate the implications of this kind of strategy, we calculate the expected number of individuals that are considered infected. First, this number includes all individuals in the subpopulations one and three. Furthermore, each 1SG(\(33\)) test in subpopulation four has a probability of \(1-(1-0.029)^{33} \approx 0.621\) to result in a positive outcome. Thus, an additional expected number of \(2\,124\,712\) individuals would be considered infected in this subpopulation. In total, an expected number of \(2\,228\,333\) individuals would be considered infected. We see that the decrease in expected cost is very limited with such a low number of tests, even if an optimized strategy is used. In particular, our lower bound shows that no strategy can reduce the expected cost below \(0.609\) using the assumed \(103\,621\) tests. However, our theory can also be used to estimate how many tests are required to achieve a certain reduction in expected cost. For example, we can find the number of tests necessary to obtain half the expected cost than without testing, i.e., \(0.478\). Our lower bound shows that at least \(0.0226\) [TpI]{acronym-label="TpI" acronym-form="singular+short"} (\(201\,256\) tests in total) are necessary, while our optimized strategies show that \(0.0419\) [TpI]{acronym-label="TpI" acronym-form="singular+short"} (\(373\,636\) tests in total) are sufficient. Our optimized strategies suggest to use a 1SG(\(4\)) strategy to test all individuals in subpopulation one and a mix of 1SG(\(24\)) and 1SG(\(23\)) strategies to test all individuals in subpopulation four. We also want to point out that an astonishing number of \(0.499\) [TpI]{acronym-label="TpI" acronym-form="singular+short"} (\(4\,447\,461\) tests in total) would be required to achieve the same goal with individual testing; and even the previously best known binary splitting testing strategies require \(0.102\) [TpI]{acronym-label="TpI" acronym-form="singular+short"} (\(909\,637\) tests in total). A second example case is presented in [\[sec:second-example-case\]](#sec:second-example-case){reference-type="ref" reference="sec:second-example-case"}, analyzing the situation in Austria during the SARS-CoV-2 pandemic in early April 2020. The two examples showcase how our theory can be used to explore possible strategies for a given set of parameters and a given number of tests. Furthermore, they already illustrate that the specific scenario can affect all parts of the optimal testing strategy, e.g., there is no general rule that the testing of individuals with highest prevalence has priority. Maybe surprisingly, the examples also illustrate that there are situations where our approach can result in substantial gains using very simple testing strategies. Any other parameter choices can be explored using the Jupyter notebook and Python code, available at <https://github.com/g-pichler/group-testing>. This software was created using a Jupyter Notebook , as well as SciPy , NumPy , and Matplotlib . Finally, we want to emphasize, that the historic example cases discussed here certainly do not completely reflect the reality of testing in Austria in 2020. In addition to the somewhat arbitrary choices of values \(b, c\), our analysis suffers from the shortcomings that are to be discussed in [6](#sec:discussion){reference-type="ref" reference="sec:discussion"}. E.g., considering the first example case ([\[fig:subpop\]](#fig:subpop){reference-type="ref" reference="fig:subpop"}), a 1SG(\(33\)) test is positive with probability around \(0.621\), while the chance of infection is less than \(0.03\). One would not expect this strategy to be viable, as compliance will be low.
# Information theoretic details {#sec:itdetails}
Readers familiar with information theory will see the obvious analogues between our problem formulation and classical rate-distortion theory. Indeed, an \((N,K)\)-test function can be seen as a specific encoder of a binary sequence, whereas the corresponding \((K,N)\)-decision procedure is a specific decoder. Thus, \(D^{(\mathbf{X}, \rho)}(K)\) defined in [\[eq:defdfink\]](#eq:defdfink){reference-type="ref" reference="eq:defdfink"} is the minimal expected distortion of a restricted class of source codes of rate \(R = K/N\) for the vector \(\mathbf{X}\). A fundamental information-theoretic result provides a lower bound on the minimal expected distortion of an arbitrary source code of rate \(R\), the so-called information distortion-rate function.
Because \(D_{\textrm{I}}^{(\mathbf{X}, \rho)}(R)\) is a lower bound on the minimal expected distortion of an arbitrary source code of rate \(R\), it can be used to obtain a lower bound on \(D^{(\mathbf{X}, \rho)}(K)\). The following is a direct consequence of Theorem 3.2.1 in and a proof is thus omitted.
To prove our main results, it remains to characterize the information distortion-rate function \(D_{\textrm{I}}^{(\mathbf{X}, \rho)}(R)\). We first focus on the case \(N=1\) and will subsequently extend our setting to prove [\[th:partitions\]](#th:partitions){reference-type="ref" reference="th:partitions"}. The following result is based in the variational description of the information distortion-rate function (see Section 2.4 in ) and a detailed proof is provided in [\[app:proofiid_rd_final\]](#app:proofiid_rd_final){reference-type="ref" reference="app:proofiid_rd_final"}.
We can now state the central characterization result of the information distortion-rate function in the setting of [\[th:partitions\]](#th:partitions){reference-type="ref" reference="th:partitions"}. A detailed proof of the result is provided in [\[app:proofsubpop2\]](#app:proofsubpop2){reference-type="ref" reference="app:proofsubpop2"}.
Combining the previous results, we can now prove [\[th:partitions,thm:RD_function\]](#th:partitions,thm:RD_function){reference-type="ref" reference="th:partitions,thm:RD_function"}.
Finally, we want to point out that many information theoretic questions about the problem remain open. In particular, the characterization of a "group testing distortion-rate function" that can be defined as the infimum of all expected distortions that are achievable at a given rate as the population size \(N\) goes to infinity is a challenging new problem that can also be formulated as a problem in classical rate-distortion theory with significant restrictions on the encoder. Here, non-adaptive group testing may be easier to analyze because the encoder is essentially restricted to the \(\max(\,\cdot\,)\) function. Another direction for future work is to incorporate testing errors, resulting in a joint source-channel coding scenario. It is not clear whether the source--channel separation theorem holds in this case. Nevertheless, we expect that a lower bound similar to the one presented here can be obtained, by adding the additional bits (i.e., tests) necessary for error-free communication over a binary (most likely asymmetric) channel.
# Discussion {#sec:discussion}
We introduced a rigorous mathematical formulation of optimal testing strategies for a given number of available group tests. The problem is formulated as a one-time testing procedure in the sense that we assume that the infection status of the population does not change during the testing procedure and we do not consider any time evolution. This enables us to use only very few parameters that describe a current pandemic situation and we do not require any detailed personal information such as contact maps between individuals. However, even these few parameters can be difficult to estimate and may also result in ethical challenges. The prevalence \(p\) is the easiest and most obvious parameter and can be estimated using a negligible number of tests for a pilot study and the estimate can be improved as the testing strategy is applied. The costs \(b\) and \(c\) of wrong assignments are more difficult to choose. They can be adapted to the infectiousness of the disease, the cost and effectiveness of quarantine, non-pharmaceutical interventions that reduce the risk of infection, and most likely many more variables. How to exactly choose these costs is beyond the scope of our work and very specific for a given situation. It may also include political decisions by weighing health factors (e.g., minimize the number of infections by quarantining many individuals) against economic factors (e.g., minimize the number of quarantined individuals). Note however that at least implicitly these costs are used in political decisions: A society-wide lockdown can be interpreted as the assertion, that assigning all (untested) individuals an infected status is less costly than considering them to be healthy. Thus, at a prevalence \(p\), this implies \((1-p) b \le p c\). Once the parameters are fixed for a given scenario, our theory on the one hand gives ultimate bounds on how large a cost reduction can be achieved by group testing, and on the other hand suggests optimal allocation of resources for simple (suboptimal) testing strategies. The present work is only a first step towards establishing a mathematical theory of group testing. Thus, we do not consider the most general scenario and focus on basic scenarios that are simplified in many aspects. In particular, the scope of our results is limited by the following assumptions:
- We assume that tests are perfect, i.e., a group test of \(u\) individuals is negative if and only if all \(u\) individuals are healthy. However, we expect that a small error probability will not significantly influence the results and simple simulations incorporating these errors can be used to check the robustness in a specific scenario. A rigorous theoretical treatment on the other hand should not only consider a fixed error probability for a group test, but the error probability should rather depend on the group size and the number of infected individuals within the group (cf. ). Thus, an extension of our work in this direction would imply a more quantitative approach of testing outcomes and not merely the binary options we consider here.
- In this work, no upper bound on the group size is assumed a priori. The information theoretic lower bounds would not be affected by such a limitation, but certain points of the \(k\)SG strategy will become infeasible, requiring minor modifications to the published code. However, a more thorough analysis would not impose a hard limit on the group size, but consider the trade-off between test accuracy and group size, as mentioned in the previous point.
- We assume that the infection status of different individuals is independent.
- We assume that the choice of testing strategy does not alter the cost of wrong assignment. This assumption may be violated in reality, as e.g., individuals that are assigned an "infected" will likely show different levels of compliance, depending on whether they were tested individually, in a group, or not at all. Going beyond this assumption could be achieved by changing the infection status from a binary decision to a probability assignment and using a suitably adapted cost function.
- The cost function is fixed and applied independently to each individual. E.g., we cannot express the fact, that quarantining a small number of individuals working in critical areas is hardly problematic but once a critical threshold is passed the cost of quarantining further individuals becomes more costly than the linear increase assumed in our model.
- We do not consider the collection of samples from individuals as a limiting factor, but merely the number of tests is limited.
- The testing strategies presented in this work are not at all optimal but merely illustrate the potential of our approach. For example, we do not consider testing individuals in several groups at the first stage (as, e.g., in the array testing approach by ), nor mixing individuals from different subpopulations within a group. Nevertheless, our suggested strategies are surprisingly simple and can outperform even the best (significantly more complicated) strategies currently known.
- We consider testing as a stand-alone task and do not incorporate it into a larger disease model, as is done, e.g., in. In particular, the probabilities of infection and the cost functions are assumed to be fixed and do not change over time.
From a theoretical viewpoint, our analysis reveals several surprising insights that are in contrast to long established fundamental statements in group testing theory. First, an established statement is that group testing is beneficial exactly in the regime \(p< \frac{1}{2}(3-\sqrt{5}) \approx 0.381\), i.e., in this regime, group testing cannot outperform individual testing. However, this result was proven under the assumption that perfect identification of the infection status of each individual is required. Maybe surprisingly, this statement does not extend to our setting. Indeed, for the specific case of \(p= \frac{1}{2}(3-\sqrt{5})\), \(b=1\), and \(c=10\), using the testing strategy 1SG(\(2\)) has a strictly lower expected cost than individually testing every second individual (both have rate \(1/2\) [TpI]{acronym-label="TpI" acronym-form="singular+short"}). Similarly, in the non-adaptive regime (i.e., all tests are performed in parallel) an established result is that individual testing is optimal if all individuals are independent and have a fixed prevalence \(p\). Again, this result only holds if perfect identification of the infection status of each individual is required. In our setting, the simple 1SG testing strategy is a non-adaptive strategy and is clearly superior to individual testing (see [\[fig:rda50\]](#fig:rda50){reference-type="ref" reference="fig:rda50"}). | {'timestamp': '2021-10-29T02:17:47', 'yymm': '2109', 'arxiv_id': '2109.05378', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05378'} |
# Introduction
Multi-Agent Path Finding is the problem of arranging a set of collision-free paths for a set of agents on a given graph. Although MAPF is NP-hard to solve optimally on graphs and even 2D grids , many optimal MAPF algorithms have been developed in recent years. Some reduce MAPF to other well-studied problems, e.g., ILP and SAT , others solve it with search-based algorithms, e.g., enhanced A\* search , Conflict-Based Search and its improved variants . However, the limitation of these centralized planning methods is that they do not scale well to a large number of agents. RL and IL methods with decentralized execution have been applied to address this issue . During execution, each agent takes action based on its individual decision model with its own observation (may also include messages from other agents) as input, which avoids the scalability problem. For RL-based methods , MAPF is generally modeled as a Markov Game with partial observability, where a reward function is designed for each agent, and the goal of each agent is to maximize its expected total return. Usually, expert guidance is applied to guide the RL during training in terms of behavior cloning , shaped reward , or heuristic map . For IL-based methods , partial observability is still considered, but the objective is to minimize the divergence between expert demonstrations and generated trajectories, so that the learned policy can mimic the expert behavior. Recently, researchers focus on trainable communication channels between agents for MAPF, where extra information can be obtained during both training and execution to enhance multi-agent cooperation . However, these methods focus on broadcast communication in which messages will be transmitted to all other or predefined set of agents. For instance, in and, the messages are broadcast to all other agents within a distance (communication radius). In, the central agent takes the messages from the two nearest neighbors after getting all the broadcast messages from its neighbors inside the field of view (FOV). Although empirical results have shown great improvements of communication based methods compared with non communication ones , the drawback is that broadcast communication requires lots of bandwidth and incurs additional system overhead and latency in practice. More importantly, not every agent provides useful information for cooperation, and redundant information can even impair the learning process. Plenty of methods on reducing communication overhead or learning selective communication have been proposed in the literature of RL . But most of them are designed for centralized training framework on cooperative Markov Game, where agents share the same team reward. In contrast, each agent has its individual reward in MAPF, and thus more information needs to be swapped for cooperation. Therefore, selective communication is much more important in this case. To design a succinct communication protocol for MAPF, we consider a request-reply communication scenario where messages are updated in two stages, and propose *Decision Causal Communication*, a simple yet efficient mechanism to enable agents to select other members in the swarm for communication during both training and execution. More specifically, before communication, each agent is capable to figure out which neighbors in its FOV are relevant and influential by just using local observation. The agent makes two temporary decisions based on its raw local observation and the modified local observation by masking out a neighbor. For any neighbor that can cause difference between these two temporary decisions, that neighbor is considered as relevant and influential, and the agent will send a request to that neighbor for communication purpose. This communication protocol is also naturally capable for decentralized execution. Empirical results show the high success rate with low communication overhead of our method compared with its counterparts. **Contributions.** In this work, we propose DCC for MAPF. Unlike broadcast communication, DCC learns selective communication, and hence encourages agents to focus only on relevant information. We demonstrate that, on one hand, selective communication can remove temporal redundant messages, which is beneficial to multi-agent cooperation. On the other hand, communication overhead is greatly cut down due to the reduction of communication frequency.
# Related work
## Reinforcement Learning based MAPF {#subsec:rl-mapf}
RL-based planners generally formulate MAPF as a multi-agent reinforcement learning (MARL) problem to learn decentralized policies for agents (robots). Compared with centralized planning methods, learning decentralized polices scales well to a large number of agents. Collision free policies are usually learned by guiding RL with constraints or demonstration data . Incorporating demonstration guidance with RL can be divided into two categories, using demonstrations generated by a centralized planner, or by a single-agent planner. For instance, a well known framework named PRIMAL builds on the asynchronous advantage actor critic (A3C) network as its RL part and uses behaviour cloning to supervise the training of RL. So it requires demonstrations generated by a centralized planner named ODrM\* . The limitation of using a centralized planner is that it requires solving a MAPF problem and is thus time-consuming, especially in a complex environment with a large number of agents. For methods using single-agent planner as guidance, MAPPER and Globally Guided RL (G2RL) use A\* search for single-agent path generation and apply an off-route penalty if agents fail to follow the path. The potential issue is that, as single-agent shortest paths are usually not unique and not globally optimal for multi-agent environment, forcing agents to follow these paths by extra shaped rewards can mislead agents. To tackle the above issues, our previous method DHC does not require a centralized planner. Although single-agent shortest paths are still adopted as guidance, DHC embeds all the potential choices of shortest paths as heuristic into the input of the model, instead of forcing agents to follow a specific path. Thus, no special shaped rewards are required. Even though empirical results have shown that enhancing RL with guidance can help to learn collision free policies for MAPF, the cooperation among agents is not directly modeled. Communication via graph convolution is a promising way to achieve multi-agent cooperation, where an agent aggregates information from its neighbors, including itself, and makes decisions based on this augmented information . This idea is recently deployed for MAPF. By treating each agent in the environment as a node and connecting neighboring nodes (agents) if they are inside the FOV of each other, a graph is formulated by DHC and multi-head dot-product attention serves as the convolutional kernel to compute interactions among agents. Similar ideas are applied by the work described in and Message-Aware Graph Attention neTwork (MAGAT) , where the communication part is a graph neural network (GNN) . However, it is worth noting that both and are IL-based methods, which train the model to imitate the demonstrations generated by Conflict-Based Search (CBS) and Enhanced CBS , respectively. Thus, MAPF instances still need to be pre-solved while extreme scenarios may fail to be solved and collected into the training data. Also, none of these communication based methods consider the efficiency of communication, resulting in high communication overhead and latency.
## Efficient and Selective Communication for MARL
Most existing works on communication in MARL, including those described above, focus on broadcast communication, i.e., broadcasting the messages to all other or predefined agents. To improve efficiency, a number of strategies have been proposed. Individualized Controlled Continuous Communication Model (IC3Net) considers that full communication is not always necessary and agents can determine whether to send messages to others indicated by a gating function. However, this gating function can block all communication channels of an agent. Cases where an agent's message is instructive to some agents but useless or harmful to others can not be handled. Schedule Communication (SchedNet) only allows a limited number of agents to broadcast messages, who are chosen by some importance weights assigned. The shortcoming is that a central scheduler is required to gather all individual weights and decide which agents should be entitled to broadcast their messages. Variance Based Control (VBC) lets agents transmit their messages only when the variance of the message vector is high (high variance implies that the message is informative). And agents decide whether to request messages based on local ambiguity. Temporal Message Control (TMC) stores agent messages in the buffer and new messages are sent out only when they contain relatively new information compared with old ones. Nearly Decomposable Q-functions (NDQ) optimizes communication via minimizing the entropy of messages between agents. Individually Inferred Communication (I2C) learns a prior network, which takes local observation and identity of another agent as input, to predict a belief of whether to communicate. All VBC, TMC, NDQ and I2C are designed on the centralized training methods that factorize the joint action-value function, such as QMIX , different from MAPF settings.
# Problem formulation {#sec:env}
## MAPF Problem Definition
MAPF has many variants as summarized in . In this paper, we consider the classical MAPF case defined in that (a) each agent performs an action in every time step, and may cause vertex and swapping conflicts, (b) agents \"stay at target\" until all agents have reached their goals, and (c) optimizes the sum of cost. This definition differs from the environments used in PRIMAL and MAPPER, which are much simpler. In PRIMAL, agents take actions in turn (only one agent takes action at a time), although we can treat this as all agents act simultaneously but only one agent can move while all others choose no-op action. In MAPPER, agents are removed from the environment upon reaching their goal locations, which simplifies the problem, as there will be no livelocks when an agent reaches its goal but obstructs other agents from getting to their goals. We also notice that the environment in is specially designed where obstacles compose narrow corridors that only allow one agent to pass at a time. We do not consider an environment like this. Formally, we define MAPF as follows. Given an undirected graph \(G=(V,E)\) and a set of \(n\) agents, each agent indexed by \(i\) has a unique start vertex \(s_i\in V\) and a unique goal vertex \(g_i\in V\). Time is discretized into time steps. At each discrete time step, each agent can either *move* to an adjacent vertex or *wait* at its current vertex. A path for \(i\)-th agent is a sequence of adjacent (indicating a moving) or identical (indicating a waiting) vertices starting at \(s_i\) and terminating at \(g_i\). Agents remain at their goal vertices (\(g_i\)) after they complete their paths. There are two kinds of collisions, either vertex collision or edge collision. A vertex collision is a tuple \(\langle i,j,v,t \rangle\) where \(i\)-th and \(j\)-th agents reaching at the same vertex \(v\) at time \(t\). An edge collision is a tuple \(\langle i, j, u, v, t\rangle\) where \(i\)-th and \(j\)-th agents traverse the same edge \((u,v)\) in opposite directions at time \(t\). A solution to MAPF is a set of collision-free paths, one for each agent. The quality of a solution is measured by the sum of arrival time of all agents at their goal vertices.
## Environment Setup
In line with standard MAPF tasks, we focus on 2D 4-neighbor grids where agents, goals, and obstacles occupy one grid cell respectively. Formally, the map is a \(m\times m\) matrix, where \(0\) represents a free location and \(1\) is an obstacle. Each map is chosen \(n\) start positions and \(n\) corresponding goal positions for \(n\) agents from free locations. We make sure each goal is reachable from its start point and there is no overlap among \(2n\) selected positions. At each time step, agents move simultaneously to neighboring locations or wait at their current locations. Thus, the size of the action space is 5 (move to four directions or stay still). Agents may hit obstacles or conflict with others during simulation. We handle the collision by recursively resetting the related agents to previous states until no collision exits. We consider a partially observable environment, which is more compatible with the real world problem. Each agent can only access the information inside its FOV with size \(\ell\times\ell\) (\(\ell<m\)), where \(\ell\) is an odd number to make sure agents are at the center of FOV. The reward design is adopted from DHC , shown in Table [\[table:reward\]](#table:reward){reference-type="ref" reference="table:reward"}. Generally, negative rewards are given to agents for each movement to facilitate goal reaching. No shaped reward is required in our method.
# Selective communication for MAPF
In this paper, we propose DCC [^1], a simple yet efficient mechanism to reduce communication overhead in multi-agent systems. DCC can be instantiated by independent Q-learning , or any framework of centralized training and decentralized execution, such as QMIX and VDN . We consider a request-reply scenario different from the traditional setup where each agent only sends out an almost information-less indicator, e.g. a scalar value, as the request signal. Instead, in our setup, each agent sends out a request signal with rich information including its own messages along with the relative positions of neighbors. In this way, after receiving the request, the agents being requested can immediately benefit from this query by collecting some information from the query agent. Fig. [\[fig:req-rep\]](#fig:req-rep){reference-type="ref" reference="fig:req-rep"} shows the difference between broadcast and our request-reply based selective communication. In the following part, We first introduce DCC communication mechanism and then instantiate DCC with independent Q-learning.
## Decision Causal Communication {#subsec:DCC}
Our communication mechanism is motivated by I2C , where causal inference is utilized to derive a threshold to trigger communication. Specifically, I2C defines the causal effect of agent \(j\) to agent \(i\) as the KL-divergence between two decision probabilities, \(D_{KL}( P(a_i|\boldsymbol{a}_{-i}, \boldsymbol{o}) || P(a_i|\boldsymbol{a}_{-ij}, \boldsymbol{o}) )\), where \(\boldsymbol{o}\) is the joint observations, \(\boldsymbol{a}_{-i}\) denotes the joint actions except for agent \(i\), and similar meaning for \(\boldsymbol{a}_{-ij}\). The magnitude of this divergence indicates how much agent \(i\) will adjust its policy if taken into account agent \(j\)'s policy. If the divergence is small, agent \(i\) chooses not to communicate with \(j\) as that will hardly affect its own policy. However, I2C is only effective for centralized training algorithms as it requires to know the joint observations and actions of all other agents. In real world problems such as MAPF, usually an agent with partial observability can only observe the existence of other agents instead of their policies. Although agents can broadcast their actions, that will definitely increase the communication overhead. In practice, if the policy of an agent requesting communication conditions on the action of the communicated agent, the circular dependencies can occur. At last, setting a reasonable threshold for that KL-divergence highly depends on empirical experiments and may vary in different problems. Thus, to design an appropriate communication protocol for MAPF, in this work, we restrict the strategy in I2C by triggering communication between agent \(i\) and \(j\) only when the *existence* of agent \(j\) causes the policy adjustment on agent \(i\), and we call this *Decision Causal Communication*. To get an intuition of the proposed protocol, consider the learning based MAPF methods discussed in Section [2.1](#subsec:rl-mapf){reference-type="ref" reference="subsec:rl-mapf"}, where most of them guide the RL agent with global or local optimal demonstrations. In order to diminish the communication cost, an agent should follow the demonstration guided policy as often as possible without communicating with others, if the existence of other agents in its FOV will not affect its current policy. Otherwise, the agent communicates with those neighbors who will individually affect its policy. Although the policy may be adjusted after communication and we can not guarantee the adjusted policy is always better than unchanged policy during simulation, training the communication part along with RL in an end to end manner to maximize the expected total return will force the communication to be helpful and performed only when it is necessary. Formally, this communication protocol can be formulated as follows. Suppose agent \(j\) can be seen by agent \(i\) (agent \(j\) is inside the FOV of agent \(i\)), and the relative position of agent \(j\) in \(i\)'s FOV is \((x_j, y_j)\). Agent \(i\) gets its local observation \(o_i\) from the environment and constructs another modified observation \(o_{i,-j}\) (\(-j\) means without \(j\)) by setting the information at agent \(j\)'s position by some special value, such as zero: \(o_{i,-j}\leftarrow o_{i}(x_j, y_j)=0\). Then agent \(i\) uses its local decision model, without communication, to predict two temporary actions \(\widetilde{a}_i\) and \(\widetilde{a}_{i,-j}\) based on \(o_i\) and \(o_{i,-j}\), respectively. If these two actions match with each other, implying that the existence of agent \(j\) will not affect \(i\)'s policy, agent \(i\) will not request message from agent \(j\). Otherwise, agent \(i\) sends a request to agent \(j\) via a communication channel, and agent \(j\) will respond with a message. In this work, we consider messages are updated in both request and reply stages, which is discussed in the following subsection. The final aggregated messages are further used by policy or Q-network to generate the final actions.
## Deep Q-Learning Model Design
We instantiate DCC with independent Q-learning, namely the model is designed and trained from a single agent's perspective by treating other agents as part of the environment. We borrow some architecture design from DHC. The whole model of DCC consists of four main components, including observation encoder, decision causal unit, communication block, and Q-network. The model architecture is shown in Fig. [\[fig:model\]](#fig:model){reference-type="ref" reference="fig:model"}. We begin from the model input and describe the four components one by one.
#### Local observation input
At time step \(t\), agent \(i\) gets a matrix with the shape \(\ell\times \ell\times 6\) as input, denoted by \(o_i^t\). The first channel is a binary matrix representing the obstacles inside the FOV. The second channel is a binary matrix indicating the locations of other agents if within the FOV. The other four are heuristic channels proposed by DHC , to encode multiple choices of single-agent shortest path. We refer readers to the DHC paper for further details.
#### Observation encoder
The encoder contains *four* stacked convolutional layers with a GRU . The local observation \(o^t_i\) is first encoded into \(\hat{o}_i^t\) by four convolutional layers. Then by adopting the last step communication outcome \(e^{[2]t-1}_i\) as the hidden state, the GRU gets \(\hat{o}_i^t\) and generates the intermediate message \(e^t_i\).
#### Decision causal unit
This unit depends on the observation encoder to embed modified observations, and the Q-network (will be discussed later), to get temporary actions. Let \(\mathbb{B}_i\) denote all neighbors of agent \(i\). Given \(o^t_i\), \(o^t_{i,-j}\) is constructed according to Section [4.1](#subsec:DCC){reference-type="ref" reference="subsec:DCC"} for all \(j\in\mathbb{B}_i\). Each \(o^t_{i,-j}\) is passed through the observation encoder as discussed above to get an embedding \(e^t_{i,-j}\). By skipping the communication, \(e^t_i\) and \(\{e^t_{i,-j}\}_{j\in\mathbb{B}_i}\) are directly fed into the Q-network to compute Q-values. Actions \(\widetilde{a}^t_i\) and \(\{\widetilde{a}^t_{i,-j}\}_{j\in\mathbb{B}_i}\) are inferred by applying argmax over Q-values. Based on these actions, the communication scope of agent \(i\) is \[\begin{aligned}
\mathbb{C}_i = \{j|\widetilde{a}^t_i\neq \widetilde{a}^t_{i,-j}\}_{j\in\mathbb{B}_i}. \end{aligned}\] Note that temporary actions \(\widetilde{a}^t_i\) and \(\{\widetilde{a}^t_{i,-j}\}_{j\in\mathbb{B}_i}\) are only used to decide the communication scope, not the final action to be executed.
#### Communication block
This block is the graph convolution with multi-head dot-production as the convolutional kernel followed by a GRU. We regard the request-reply as a two-round communication. After defining the communication scope \(\mathbb{C}_i\), at the first round, agent \(i\) sends request information to all agent \(j\in\mathbb{C}_i\). This information includes its own message \(e^t_i\) as well as the relative position of \(i\)'s neighbors inside \(i\)'s FOV, denoted as \(l^t_i\). The relative position of each neighbor is originally represented as a one-hot vector, i.e. the vector length is \(\ell\times\ell\), and further embedded by a neural network layer. At the first round, agent \(j\) may receive many requests from different agents, and we define this request receiving scope for agent \(j\) as \[\begin{aligned}
\bar{\mathbb{C}}_j = \{i|j\in \mathbb{C}_i\}. \end{aligned}\] Then each agent \(j\) who receives requests will update its message using multi-head attention. Let \(\bar{\mathbb{C}}_{j+}\) denote the set \(\{j,\bar{\mathbb{C}}_j\}\). For every agent \(\bar{i}\in\bar{\mathbb{C}}_{j+}\), agent \(\bar{i}\)'s own message \(e^t_{\bar{i}}\) is projected to **Q**uery by matrix \(\boldsymbol{W}^h_Q\), while the concatenation of \(e_{\bar{i}}^t\) and \(l_{\bar{i}}^t\) is projected to **K**ey and **V**alue by matrix \(\boldsymbol{W}^h_K\) and \(\boldsymbol{W}^h_V\) in each independent attention head \(h\). The relation between agent \(j\) and \(\bar{i}\in\bar{\mathbb{C}}_{j+}\) in \(h\)-th attention head is computed as \[\begin{aligned}
\mu^h_{j\bar{i}}=\mathrm{softmax}\left(\frac{\boldsymbol{W}^h_Q e^t_j\cdot (\boldsymbol{W}^h_K [e^t_{\bar{i}},l^t_{\bar{i}}])^{\top} }{\sqrt{d_K}}\right), \end{aligned}\] where \(d_K\) is the dimension of **K**eys. The **V**alues are weighted summed by weights \(\mu^h_{j\bar{i}}\) over \(\bar{\mathbb{C}}_{j+}\) at each head \(h\). And all head outputs are concatenated over \(\mathcal{H}\) heads to pass through a neural network layer \(f_o\) for the final output \[\begin{aligned}
\hat{e}^t_j=f_o\left(\mathrm{concat}\left(\sum_{\bar{i}\in\bar{\mathbb{C}}_{j+}} \mu^h_{j\bar{i}}\boldsymbol{W}^h_V[e^t_{\bar{i}},l^t_{\bar{i}}], \forall h\in\mathcal{H}\right)\right). \end{aligned}\] The message \(\hat{e}^t_j\) and the initial message \(e^t_j\) are first aggregated by a GRU to generate the first round outcome \(e^{[1]t}_{j}\). Then \(e^{[1]t}_{j}\) acts as the initial message for the next round. At the second round, agents who received requests will reply with their updated messages along with the relative position of their neighbors, i.e. agent \(i\) will receive \(e^{[1]t}_j\) and \(l_j^t\) from all agent \(j\in\mathbb{C}_i\). Let \(\mathbb{C}_{i+}\) denote the set \(\{i,\mathbb{C}_i\}\). Agent \(i\) will update its message in the same manner as defined in Equations 3-4 by replacing \(j\) with \(i\), and replacing \(\bar{i}\in\bar{\mathbb{C}}_{j+}\) with \(\bar{j}\in\mathbb{C}_{i+}\). Denote the message generated by Equations 3-4 at this round as \(\hat{e}^{[1]t}_i\). Finally, \(\hat{e}^{[1]t}_i\) and \(e^{[1]t}_j\) are aggregated by a GRU to output \(e^{[2]t}_i\), the final outcome message of the second round.
#### Q-network
The message \(e^{ [2]t}_i\) is adopted by a dueling Q-network , which separates state advantage \(V_s(\cdot)\) and action advantage \(A(\cdot)\), to predict the Q-values. \[\begin{aligned}
Q_{s,a}^i=V_s(e^{[2]t}_i)+\left(A(e^{[2]t}_i)_a-\frac{1}{|\mathcal{A}|}\sum_{a'}A(e^{[2]t}_i)_{a'}\right), \end{aligned}\] where \(\mathcal{A}\) is the action space. After we get the \(Q\) values of agent \(i\), the loss function for training the whole model is a time difference (TD) error. To stabilize training, we use a multi-step TD error calculated by the mean square error (MSE) \[\begin{aligned}
\mathcal{L}(\theta)=\mathrm{MSE}\left(R^i_t-Q^i_{s_t,a_t}(\theta)\right) \end{aligned}\] with \(R^i_t=r^i_t+\gamma r^i_{t+1}+...+\gamma^{n}Q^i_{s_{t+n},a_{t+n}}(\bar{\theta})\), where \(r^i_t\) is the reward received by agent \(i\) at time \(t\), \(R^i_t\) is the multi-step expected return, \(\theta\) represents the parameters of the entail model, and \(\bar{\theta}\) denotes the parameters of the target network, a periodical copy of the online parameters \(\theta\).
## Training {#subsec:training}
As mentioned above, we train the model from a single agent's perspective using independent Q-learning, where each agent learns its own action value \(Q^i\) independently and simultaneously, treating other agents as part of the environment. The benefit is that it avoids the scalability problem in centralized training, which requires learning a Q-function for joint actions over all agents, as the joint action space grows exponentially when the number of agents increases. To further simplify the training process, instead of training separate models for different agents, we train a single model by sharing parameters among agents. To facilitate training, we adopt two training strategies, namely distributed training and curriculum learning , which are discussed in the following.
#### Distributed training
Distributed training has significantly improved the performance of RL compared with non distributed ones, such as R2D2 and Ape-X . In MAPF, PRIMAL is a distributed RL method by using A3C. The idea behind this distributed version of algorithm is to parallelize the gradient computation, to more rapidly optimize the parameters. Another approach, as proposed in Ape-X, is to parallelize experience data generation and selection with a shared replay memory. We adopt the latter strategy, where we setup multiple runners on CPUs with their own copy of the environment and up to date model to generate experience data, and a single learner on GPU to train. The runners initialize the priorities for the experience data and feed into a global prioritized reply buffer. The learner samples the most useful experiences from the buffer to update the model parameters and also update the priorities of the experiences. The advantage is that the model training is allocated to GPU while original A3C is designed for multi-core CPU. Note that the transitions of all the agents need to be recorded for communication purpose in the model. As the priorities are shared among all runners, the good experiences explored by any runner can benefit the learner. As we only have a single learner on GPU, the core task is to efficiently distribute multiple runners on CPUs. Here we utilize a powerful Python package named Ray , a distributed framework designed for machine learning, to easily deploy parallel programs on multiple CPUs with little modification to the deep learning code.
#### Curriculum learning
The ultimate goal of the learning model is to handle complex cases in a large environment with high obstacle density and many agents. However, it would be hard for the model to learn fast and stably if directly starting from a complicated training environment. As pointed in , presenting training examples not randomly but in a meaningful order will benefit machine learning algorithms. Motivated by this idea, we design a training pipeline to gradually introduce more complex training cases to the model. Specifically, training starts with a task where only \(1\) agent stays in a \(10\times 10\) environment. Then if the success rate of the current task exceeds \(0.9\), two new tasks are established for training. One is to increase the number of agents by \(1\), and the other is to increase the size of the environment by \(5\). The final task with \(16\) agents in a \(40\times 40\) environment will be reached as the training scale grows.
#### Other training settings
For the environment setting, the obstacle density for each environment during training is sampled from a triangular distribution between \(0\) and \(0.5\) with a peak at \(0.33\), which is the same as the obstacle setting in PRIMAL. The FOV size is \(9\times 9\) (\(10\times 10\) in PRIMAL, we make it odd) and the agent is in the middle of the FOV. The maximum step for the environment is \(256\). For training setting, we train the model with a batch size of \(128\) and a sequence length of \(20\). The discount factor \(\gamma=0.99\). The multi-step TD error defined in Equation 5 is computed with a length of \(2\). The number of runners for distributed training is set to be \(16\).
# Experiments
The model is trained and tested in classical MAPF environments as discussed in Section [3](#sec:env){reference-type="ref" reference="sec:env"}. We compare DCC with one of the most well known RL baselines named PRIMAL , and the most recent RL plus communication method named DHC in terms of success rate and average step. For comparing communication overhead, we build a baseline model, which uses request-reply mechanism but with predefined communication scope. PRIMAL and DHC are not included in this evaluation because they are all-to-all communication (more costly than request-reply).
## Success Rate and Average Step
Success rate measures the ability to complete a MAPF task within given time steps. Average step measures the average time step consumed to finish a task, where smaller value indicates a better policy. We average both successful and unsuccessful cases to calculate the average steps. We set up two types of maps, \(40\times 40\) and \(80\times 80\) for testing. The obstacle density is set to be \(0.3\), the highest testing density in PRIMAL and DHC. We set up 200 test cases for each agent number in \(\{4,8,16,32,64\}\). Additional agent number of \(128\) is tested in \(80\times 80\) maps due to a larger environment space. The maximum time step for \(40\times40\) map is 256, and 386 for \(80\times80\) map, the same as the PRIMAL's setting. Fig. [\[fig:res_compare\]](#fig:res_compare){reference-type="ref" reference="fig:res_compare"} shows the success rate of our method compared with PRIMAL and DHC in two types of environments. In general, DCC and DHC (RL plus communication) perform much better than PRIMAL (RL with expert guidance) in all cases. The performance of PRIMAL downgrades heavily in relatively larger environments (\(80\times 80\)) compared with smaller ones (\(40\times 40\)), which indicates the IL part of PRIMAL does not deliver good guidance to its RL component. Thus it suffers performance degradation on long-horizon task. In the environments where the agent density is low, such as \(4\), \(8\), \(16\) agents in both \(40\times 40\) and \(80\times 80\) maps, the difference between the success rate of DCC and DHC is small. In these environments, agents have a lower chance to meet and communicate with others, so the difference between broadcast communication (DHC) and selective communication (DCC) is not very significant. However, the success rate of DHC drops rapidly when agent density further grows. In agent dense environments, agents have a higher chance of encountering each other, but agents will obtain irrelevant and redundant information from others due to broadcast communication, which harms the cooperation among agents. By selecting and gathering relevant and influential information for communication, the cooperation is enhanced by DCC. Table [\[table:avg_step\]](#table:avg_step){reference-type="ref" reference="table:avg_step"} verifies that DCC learns higher quality policies with respect to average steps. We include ODrM\* , a centralized planner, as a reference. Especially compared with DHC, learning selective communication by DCC leads to much shorter paths in agent dense environments.
## Selective Communication vs. Predefined Scope
Generally, selective communication by DCC does not need to set a communication scope. Any neighboring agent inside the FOV is possible to be requested by the central agent for communication. Previous broadcast communication method DHC defines the communication scope as the nearest two neighboring agents inside the FOV in order to reduce temporal redundant information. Although it is reasonable to assume the nearest agents are most relevant and influential, we show that selective communication outperforms the predefined scope (nearest two agents) for MAPF in both success rate and communication overhead. To make a fair comparison, we develop a baseline model, named RR-N2 (request-reply, nearest two). The whole model architecture is similar to DCC. The only difference is that RR-N2 does not have the decision causal unit in Fig. [\[fig:model\]](#fig:model){reference-type="ref" reference="fig:model"}, and in communication block, each agent only requests information from its two nearest agents inside the FOV. The training settings are the same as DCC as discussed in Section [4.3](#subsec:training){reference-type="ref" reference="subsec:training"}.
Fig. [\[fig:res_self\]](#fig:res_self){reference-type="ref" reference="fig:res_self"} shows the success rate of our method compared with this baseline. The difference in success rate is significant in agent-dense environments, i.e., \(64\) agents in \(40\times 40\) map and \(128\) agents in \(80\times 80\) map, which implies the nearest two agents are not always the most relevant ones. One may think that DCC achieves a better performance due to unconstrained communication scope, where each agent may communicate with more than two agents inside its FOV and thus gathers more information than RR-N2. We show that this is not really the case and DCC actually learns succinct communication. In particular, we compute the average communication frequency in these test environments as shown in Table [\[table:comm_freq\]](#table:comm_freq){reference-type="ref" reference="table:comm_freq"}. A pair of request and reply is counted as one time communication. The average communication overhead of RR-N2 is much more costly than DCC in all the testing cases, and is extremely expensive in agent-dense environments. By learning selective communication, DCC can greatly reduce communication by only focusing on relevant messages. Combining these two results, we can conclude that although communication can help multi-agent cooperation for MAPF, agents should learn to actively select relevant agents for communication, because even the most nearest agents are not the relevant ones. Communicating with only two agents can already lead to huge communication overhead. Thus, researchers should always devise a succinct communication mechanism in order to reduce the communication overhead for easy deployment to real world problems. **Limitations:** There is an additional computational cost associated with selecting neighboring agents. In extreme cases, the cost is of order of \(O(n\times p)\). \(p\) is the FOV capacity.
# Conclusion
We propose DCC to learn succinct communication for MAPF with classical MAPF environment settings. Our aim is to enable agents to actively select relevant and influential neighbors for communication, instead of broadcasting. The selection is learned via the local policy adjustment effect between agents, which captures the necessity of communication. Empirical results show that selective communication with relevant agents improves the policy learning process. Moreover, DCC also serves as a component for communication reduction, greatly downscaling the communication overhead. Future work entails extensions of DCC to centralized training frameworks and cooperative Markov Game tasks. Another direction is to explore how to directly exclude sets of neighbors during neighboring selection to reduce complexity, and also the maximal number of agents that DCC can handle in different size of maps.
[^1]: Code available at https://github.com/ZiyuanMa/DCC | {'timestamp': '2021-12-24T02:22:13', 'yymm': '2109', 'arxiv_id': '2109.05413', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05413'} |
null | null |
null | null |
null | null |
# Introduction
Be/X-ray Binaries (BeXRBs) are high mass X-ray binaries (HMXB) consisting of typically a Neutron Star (NS) compact object and a main sequence star of B spectral type, showing strong emission line features in its optical spectrum, most commonly hydrogen Balmer lines, but also may show He and Fe emission lines . Although NS are the most common compact object type in these objects, in some rare cases the compact object in the system has been reported to be a Black Hole, or a White Dwarf (WD), most recently the WD containing BeXRB (hereafter BeWD) Swift J004427.3-734801 discovered in the SMC. BeXRBs are often discovered through two modes of X-ray emission, typically referred to as Type-I and Type-II outbursts. Type-I outbursts (\(L_\mathrm{X} \approx 10^{36}-10^{37}\) erg/s), the most common, occur when the compact object interacts with the Be star. This accretion occurs when the compact object passes through a circumstellar disc (CSD) which forms around the Be star. The presence of the CSD is strongly inferred from the presence of excess infrared emission over that expected from the star, which brightens when the CSD grows . Type-I X-ray outbursts are transient, and their onset are linked to the growth of the CSD (e.g. @Kennea2020). Type-II outbursts are typically at least an order of magnitude brighter than Type-I (\(L_\mathrm{X} \gtrsim 10^{37}\) erg/s), sometimes reaching Super-Eddington luminosities (e.g. in SMC X-3, @Townsend2017), and are not obviously linked to the orbital period of the source, with outbursts that can span multiple orbital periods. The extended length Type-II outbursts suggest the formation of an accretion disc around the compact object, but the conditions required to form this are unclear, with suggestion that warping of the CSD plays a role (e.g. @Negueruela01). When the compact object is a NS, an X-ray pulsar is typically seen, along with an X-ray spectrum which in the soft X-ray band (\(0.5-10\) keV) is well described by a hard power-law spectrum with photon indices of \(0-1\) typical. In the case of BeXRBs with WDs, the X-ray spectrum seen is more typical of a so-called Super Soft Source (SSS; e.g. ), in which the X-ray emission is thermal in nature and is predominantly emitting below 1 keV. The irregular dwarf galaxy the Small Magellanic Cloud (SMC) lies at a distance of 62.1 kpc, with a very low X-ray absorption in the line-of-sight of \(N_\mathrm{H} = 5.9 \times 10^{20}\) cm\(^{-2}\). Although many X-ray binaries are known in the SMC, all are HMXB, and predominantly they are BeXRBs, which has made it a well studied area of the sky to look for BeXRBs in the past (e.g. @haberl2000 [@Haberl2016; @CK2015]). Motivated by the study of BeXRBs, but with more of a focus on long term transient behaviour, the Swift SMC Survey (S-CUBED; @kennea2018) began in 2016. S-CUBED consists of a weekly survey of the optical extent of the SMC utilizing the Neil Gehrels Swift Observatory (Swift; @gehrels04). The survey consists of 142 pointings with an exposure of 60s each. Data are taken utilizing Swift's X-ray Telescope (XRT; @burrows05) in Photon Counting (PC) mode and Ultra-violet/Optical Telescope (UVOT; @Roming05) utilizing the *uvw1* filter. Due to the low background of XRT PC mode, 60s exposure allows for detection of an X-ray source in the SMC at a brightness of \(1-2\%\) Eddington Luminosity for a \(1.4~\mathrm{M}_\odot\) NS or brighter. In this paper we report the discovery, by S-CUBED, of a likely new BeXRB containing a WD compact object, . We present results of both S-CUBED and Swift target of opportunity (TOO) observations, as well as reporting on historical I-band light-curve from the Optical Gravitation Lensing Experiment (OGLE; @Udalski2003 [@Udalski2015], and spectroscopic observations by the South African Large Telescope (SALT).
# Observations
## Discovery of
was first discovered by the S-CUBED survey, and was reported by. The new source, internally designated SC1825 and subsequently named using the standard Swift naming scheme, was first detected in S-CUBED data taken on 2020 December 29, with an initial brightness of \(0.05^{+0.07}_{-0.03}\) count s\(^{-1}\). No previous detection of this source was found in any of the S-CUBED observations going back to when the survey started on 2016 June 8. Subsequent observations saw the brightness of the source rise to \(0.15^{+0.07}_{-0.03}\) count s\(^{-1}\) on 2021 January 5 S-CUBED observation, and then \(0.19^{+0.08}_{-0.03}\) count s\(^{-1}\) in the 2021 Jan 12 observation. At this point, a Swift Target of Opportunity (TOO) request was triggered to obtain more detailed observations. The most accurate X-ray position for , determined using the method described by, was found to be R.A.(J2000) = \(01^h 15^m 10.93^s\), Dec.(J2000) = \(-72^\circ 56' 10.4''\) with an estimated uncertainty of \(2.0''\) radius (90% confidence). In UVOT data we find a single bright star inside the XRT error circle. A catalog search of this source revealed a known star, 2MASS J01151085-7256102, with reported spectral types of O9IIIe and O8-O9Ve, suggesting that this source was a new BeXRB.
## Swift Observations
The field containing has been regularly observed as part of the weekly S-CUBED survey starting on 2016 June 8th. As of 2021 July 27, it has been observed with S-CUBED on 199 occasions, with a median exposure time of 60s per observation. Although S-CUBED observations are planned weekly, observation gaps occur frequently due to Swift visibility or scheduling constraints. In addition to the S-CUBED observations, detection of the outburst lead to TOO observations of approximately every 1-2 days, starting on 2021 January 13 with the final exposure here taken on 2021 May 11. In addition, in 2021 July, late time observations totaling 5.08 ks exposure were taken post-outburst in order to determine a strong upper limit on quiescent emission from . A detailed list of TOO observations is given in Table [1](#tab:x1){reference-type="ref" reference="tab:x1"} and the measurements are shown in Fig. [\[fig:xlc\]](#fig:xlc){reference-type="ref" reference="fig:xlc"}. For all XRT observations the data were collected in PC mode. UVOT data were collected in *uvw1* for S-CUBED data, and in the all three UV filters for TOO observations. In addition some of the TOO observations also collected data in \(u\), \(b\) and \(v\) filters. Although the TOO observations aimed for twice daily cadence, due to scheduling and spacecraft constraints, there are gaps in the coverage, especially during periods of time when the source was near the Swift orbit pole, when observations are not possible for up to 10 days.
We have analysed XRT and UVOT data from both Swift S-CUBED observations and TOO observations, in order to produce long term light-curves and pre-outburst upper limits. Observations analysis was performed utilizing the standard Swift HEAsoft v6.28 tools, including `xrtpipeline` for XRT processing and extraction of spectra, and `uvotmaghist` for UVOT photometry. Extraction of XRT light-curves, include correction of detector issues such as hot pixels, hot columns and pile-up, was performed using the Python `swifttools` module[^1], which implements the methods presented by. Spectral fitting was performed utilizing XSpec. The outburst light-curve is shown in Fig. [\[fig:xlc\]](#fig:xlc){reference-type="ref" reference="fig:xlc"}, showing a double peaked outburst, with the observed brightness peaking on 2021 January 27 (MJD 59231.8), at a luminosity of \(3.3 \pm 0.2 \times 10^{36}\) erg s\(^{-1}\) (0.5-10 keV), before fading, followed by a second period of outburst that peaked at a \(1.0 \pm 0.2 \times 10^{36}\) erg s\(^{-1}\) (0.5-10 keV) around 2021 April 13 (MJD 59317.1). From initial detection by S-CUBED to the final detection in a TOO observation taken on 2021 April 29 (MJD 59333.1), the outburst is lasted approximately 120 days. Note that all luminosity measurements assume a standard SMC distance of 62.1 kpc. Following these two outburst peaks, the source faded to non-detection, and TOO observations ceased. S-CUBED monitoring of the source continued, but no detection of was made after the final upper TOO detection limit was measured on 2021 April 29. Combined late time observations taken in 2021 July find a \(3\sigma\) upper limit of \(<2 \times 10^{34}\) erg s\(^{-1}\) (0.5-10 keV), showing that the quiescent X-ray emission was at least 2 orders of magnitude fainter than the peak brightness. We note that does not appear in any X-ray catalogue prior to discovery, despite being in fields observed by both Chandra and XMM-Newton in the past. In order to obtain a high quality X-ray spectrum, we combined all the Swift XRT TOO data. The resultant spectrum is very soft, with the majority of X-ray emission below 2 keV (see Fig. [\[fig:spectrum\]](#fig:spectrum){reference-type="ref" reference="fig:spectrum"}). This is in stark contrast to typical BeXRB in which the X-ray emission is typically a hard power-law with photon index \(\simeq1\). The very soft spectrum suggests that, analogous to the previously reported Swift J004427.3-734801, is a BeXRB with a WD compact object. The X-ray spectrum is best fit by an absorbed blackbody model (Xspec `bbodyrad`), with mean \(N_\mathrm{H} = 0.194^{+0.026}_{-0.024} \times 10^{21}\) cm\(^{-2}\) (using the Xspec `TBabs` absorption model with abundance set to `wilm`, e.g. @Wilms00). This absorption is higher than the expected level of \(5.9 \times 10^{20}\) cm\(^{-2}\), however we note that Swift J004427.3-734801 also showed enhanced absorption over the expected line-of-sight absorption, which may suggest localized absorption is a common feature in BeWD binaries. notes that excess absorption over line-of-sight is a common feature in accreting WD binaries. The fit is significantly improved (reduced \(\chi^2 = 0.95\) versus \(2.14\)) by the inclusion of an absorption edge (using the Xspec `edge` model) with a fitted edge energy \(E_\mathrm{Edge} = 0.864\pm0.011\) keV, and an edge depth \(\tau_\mathrm{Edge} = 1.84 \pm 0.32\). This is likely the signature of the O [viii]{.smallcaps} 0.871 keV absorption edge, a common feature of WD atmospheres frequently seen in SSS (e.g. @Shimura00). The fitted average black body temperature is \(kT_\mathrm{bb} = 96.7 \pm 4.2\) eV. Assuming spherical accretion on the WD, shows that for a WD with mass \(1.2 \mathrm{M}_\odot\), the maximum shock temperature would be 92 eV, consistent with the measured blackbody temperature in . Therefore the X-ray emission here is consistent with an accreting WD scenario, with the mass of the WD of \(\sim1.2~\mathrm{M}_\odot\). In order to make an estimate of the maximum radius of the emitting region, we fit the spectrum of at peak luminosity, with the same model as above, with \(N_\mathrm{H}\), \(E_\mathrm{Edge}\), \(\tau_\mathrm{Edge}\), and \(kT_\mathrm{bb}\) all fixed at the above average values, with only normalization free. The normalization is equal to \(R_\mathrm{bb}^2/D_{10}^2\) where \(R_\mathrm{bb}\) is the emission radius in kilometers, and \(D_{10}\) is the distance to the source in units of 10 kpc, allowing emission radius to be calculated if the distance is known. From this fit we find a blackbody emission radius of \(R_{\mathrm{bb}} = 1642 \pm 83\) km, assuming the standard SMC distance of 62.1 kpc. This value is order-of-magnitude consistent with the estimated \(\sim4000\)km radius of a \(1.2~\mathrm{M}_\odot\) WD (e.g. @Pringle75), consistent with the hypothesis that is a BeWD in the SMC, rather than a foreground object. Analysis of the UVOT light-curve derived from both S-CUBED data and the TOO observations do not show any significant brightness changes in the optical/UV counterpart. Fig. [\[scubed_lc\]](#scubed_lc){reference-type="ref" reference="scubed_lc"} shows the S-CUBED X-ray and UVOT *uvw1* light-curve for the entire period that S-CUBED monitoring the source, starting in 2016 June. Significantly, the outburst of was not preceded or accompanied by any UV brightening. Fitting a constant level model to the UVOT *uvw1* light-curve reveals that it is consistent with no statistically significant variability, with reduced \(\chi^2 = 0.9\). Therefore there is no detectable UV signature of this outburst.
## OGLE optical photometry
The OGLE project undertakes to provide long term I-band photometry with an average cadence of 1-3 days. The star 2MASS J01151085-7256102 was observed continuously for nearly 2 decades in the I-band until COVID-19 restrictions prevented any further observations from March 2020. It is identified in the OGLE catalogue as:\
\
O-III: smc115.1.38 I-band\
O-IV: smc732.06.6370 I-band\
OIII: smc115.1.v.17 V-band\
OIV: smc732.06.v.7898 V-band\
The I band data are shown in their entirety in Fig. [\[fig:merged\]](#fig:merged){reference-type="ref" reference="fig:merged"}. The overall behaviour revealed is consistent with that expected for a Be star in the SMC, showing large scale fluctuations on timescales of years. Also shown in this figure are the (V-I) colour changes observed by OGLE during this period. The source has clearly been much redder in colour since its return to its brightest state. This is discussed further in the Section [3](#discussion){reference-type="ref" reference="discussion"} below. Data from the most stable section of the OGLE observations, namely the 8 years after MJD 56500, were selected for a Lomb-Scargle timing analysis. The period range 2--100 d was investigated and one strong peak emerged at the period of 17.402 d. The power spectrum for the range 10-25d is shown in Fig. [\[fig:ls\]](#fig:ls){reference-type="ref" reference="fig:ls"}. The side peaks are the result of the true period beating with the annual sampling. This subset of OGLE data was then folded at the period of 17.402 d and the resulting profile is shown in Fig. [\[fig:2panels\]](#fig:2panels){reference-type="ref" reference="fig:2panels"}.Though there is some scatter in the individual cycles (as is shown in the lower panel of that figure) there clearly exists a dominant peak in the I-band light covering the phase range \(\sim\) 0.3 of a cycle, i.e. \(\sim5\) days. The ephemeris for the time of the peak of the optical outbursts, \(T_{\mathrm{opt}}\), is: \[T_{\mathrm{opt}} = 56103.7 + N(17.402)~\textrm{MJD}\label{eq:1}\]
## SALT observations
2MASS J01151085-7256102 was observed with the SALT on 2021 June 19 (JD 2459384.7) using the Robert Stobie Spectrograph. The exposure time was 300 s and the grating used was PG0900. The PG0900 grating produces a spectrum over the range 4000--7000 Å, at a resolution of approximately 1700 at H\(_{\alpha}\) and 1100 at H\(_{\gamma}\). The SALT observations of the two main hydrogen Balmer lines are shown in Fig. [\[fig:salt2\]](#fig:salt2){reference-type="ref" reference="fig:salt2"}. H\(\alpha\) is clearly shown in emission, consistent with a single peak, and with an Equivalent Width of-3.2 \(\pm\) 0.1 Å. The H\(\beta\) line is predominantly in absorption but with definite evidence of partial infilling. Such Balmer line emission is characteristic of the presence of a circumstellar disk around the B-type star. The blue end of the spectral range is shown in Fig. [\[fig:salt3\]](#fig:salt3){reference-type="ref" reference="fig:salt3"}. The spectrum shows prominent HeII lines at 4200, 4541, 4686 and 5412 Å, as well as the ionized C+O blend at 4650 Å. These lines show that the star is a late-type O star. As outlined in and discussed in, the ratio of HeII 4541 / HeI 4471 is \< 1, indicating a spectral type later than O7. The ratios HeII 4200 / HeI 4143 and HeII 4541 / HeI 4388 are both approximately 1, signifying a spectral type of O9. This is supported by the clear identification of CIII 4650 (blended with the weaker OII line) and possible detection of SiIV 4089 (though we note it is very difficult to identify this line for certain above the noise of the continuum); both indicators of late-type O stars. It is difficult to distinguish between luminosity classes III-V due to the absence of strong metal lines in the SMC. However, our spectrum matches very well with the one presented in (see the bottom spectrum of their Fig. 8). Those authors conclude an O9 IIIe classification based partly on the V-magnitude expected at the distance of the SMC. Based on our spectrum, we fully agree with their classification.
# Discussion
like many other BeXRB systems was discovered through X-ray outburst, in this case the outburst itself was prolonged, much longer than the measured 17.4 day orbital period, suggesting a Type-II outburst. The soft thermal spectrum and presence of an O [viii]{.smallcaps} absorption edge, commonly seen in WD sources such as SSS, clearly identifies the compact object in this system as a WD. The proposed optical counterpart, 2MASS J01151085-7256102, has a spectral type of O9IIIe, suggesting that is in fact a newly discovered BeWD binary system. would only be the sixth candidate BeWD identified, fourth in the SMC, following on from Swift J004427.3-734801 which was also discovered by S-CUBED, XMMU J010147.5-715550, Suzaku J0105--72, and the two Large Magellanic Cloud (LMC) BeWD candidates, XMMUJ 052016.0-692505 and RX J0527.8-6954. Fig. [\[fig:cmd\]](#fig:cmd){reference-type="ref" reference="fig:cmd"} shows the strong correlation between the brightness of 2MASS J01151085-7256102 and the redness of the whole system. The observed light is arising as a result of a fixed contribution from the OB star plus a variable contribution from the, generally cooler, circumstellar disc. Taking the spectral class of the star to be O9IIIe with a colour of (V-I) =--0.37, and assuming the reddening to the SMC to be E(V-I)=0.067, this implies that the underlying observed colour of the star, without any circumstellar disc, would be (V-I) \(\sim\)--0.30. It is clear from Fig. [\[fig:merged\]](#fig:merged){reference-type="ref" reference="fig:merged"} that the overall colour never reaches that blue extreme during the time of these OGLE observations, indicating the continual presence of a significant, but variable circumstellar disc in this system. reported that in the case of the S-CUBED discovered BeXRB Swift J004516.6\(-\)`<!-- -->`{=html}734703, the UVOT *uvw1* brightness tracked the OGLE-*I* brightness. The OGLE-*I* band light-curve in Fig. [\[fig:2panels\]](#fig:2panels){reference-type="ref" reference="fig:2panels"} taken between the June 2016 and March 2020 show that the source not showing any large scale variations during this period (excluding orbital variations). Similarly the UVOT *uvw1* light-curve (Fig. [\[scubed_lc\]](#scubed_lc){reference-type="ref" reference="scubed_lc"}) is steady during this period and after, including the period of the X-ray outburst. If *uvw1* brightness is a proxy for \(I\)-band, then this is suggestive that the optical state remained stable after March 2020. We note that no UV enhancement is expected associate with the super-soft X-ray emission, as the extrapolated UV brightness is too faint to be detected. The timings of all the XRT detections in were investigated, as was the strength of each detection. Folding both against the phase given in Eqn. [\[eq:1\]](#eq:1){reference-type="ref" reference="eq:1"} does not reveal any preferences for X-ray emission at any specific binary phase. This is, perhaps, not surprising if we are seeing a classic Type II outburst. Such events are believed to occur when the circumstellar disc expands sufficiently far to encompass the whole of the neutron star orbit. Type-I outbursts, if they exist in these BeWD binary systems, are likely below the level of detection of the shallow S-CUBED exposures. The small size of the H\(\alpha\) emission seen in this system puts it amongst a group of other High Mass X-ray Binary (HMXB) systems which also show a small Equivalent Width (\(\ge\)-10Å), but still exhibit significant X-ray activity. There are five such systems known in the SMC (SXP 8.65, SXP 8.80, SMC X-1, SXP 101, SXP 2.37) and one in the Milky Way (SAX J2103.5+4545) that have well documented small H\(\alpha\) EW values and known binary periods. Following the technique described in it is possible to estimate the size of the circumstellar disks & the size of the orbital radii, and then compare these values with that found here for . The technique relies upon estimating the circumstellar disk size from the strength of the H\(\alpha\) emission, and the size of the orbit from the orbital period and the mass of the OB star. Using that approach for the circumstellar disk size is estimated to be \((3.35\pm 0.14)\times 10^{10}\) m and the orbital size to be \((5.47\pm 0.17)\times 10^{10}\) m. This result is shown in Fig. [\[fig:disk\]](#fig:disk){reference-type="ref" reference="fig:disk"}. It is immediately apparent that a strong correlation exists between these two parameters, confirming that in these tight orbital situations the orbiting compact object plays a major role in constraining the extent of the circumstellar disc. Such restrictions have been proposed by based upon their Smooth Particle Hydrodynamic simulations. Though the compact object in this system is proposed here to be a WD rather than a NS for the other systems, the location of on this diagram strongly suggests that such an object can also play a very similar role to that of the neutron star in the other systems. However, studies of further Be white dwarf systems are needed to confirm that no significant differences actually exists. A large population of BeWDs has been predicted by binary evolution modeling, however despite this relatively few such systems are known. Swift J004427.3-734801 became the only third such system to be discovered in the SMC, despite the apparent overabundance of BeXRBs compared to the Milky Way. The fourth discovered SMC BeWD, shares similar properties to Swift J004427.3-734801, with orbital periods of \(\sim20\) days, and both being discovered through apparent Type-II X-ray outbursts. It is clear that without a sensitive, soft-band regular X-ray survey of the SMC such as S-CUBED, the likelihood of discovery of sources like and Swift J004427.3-734801, is much smaller. The peak fluxes of these sources, combined with their soft spectra put them below the level of detection of current all-sky survey instruments such as MAXI, Fermi/GBM and Swift/BAT. If Type-I outbursts are absent or too faint to be detected by S-CUBED in SMC BeWDs, to detect X-ray emission from a BeWD it needs to be undergoing a Type-II outburst, which are brighter and longer lasting than Type-I. Unfortunately Type-II outbursts are also much less frequent, with often years between outbursts. So in order to discover new BeWDs, observations need to occur in the right place, i.e. in a location where BeWDs are likely to exist and are detectable (i.e. low absorption in the line of sight), and at the right time (when they are undergoing a Type-II outburst). Given this, the relative paucity of known BeWDs can at least partially be explained due to observational bias. This highlights the importance of sensitive soft X-ray surveys as a discovery engine of BeWD binaries. It is therefore not surprising that Swift J004427.3-734801 was also detected by the all-sky soft X-ray survey telescope eROSITA. However, the focused SMC only approach of S-CUBED has proven a powerful tool for discovery of BeWD, by in 5 years doubling the known number of SMC BeWDs. It is notable that as yet, a BeWD has not been discovered in the Milky Way, only in the SMC and LMC. This may be a selection effect due to much higher extinction rates in the plane of the Milky Way, combined with the observational biases discussed above. On the other hand the different metallicity of the Magellanic Clouds may play a crucial role in the evolution of such systems. We note that BeWD are important systems, as they have been suggested as progenitors of peculiar overluminous Type Ia, and possibly also super luminous Type II supernovae such as SN 2006gy. In addition they may also play a role as progenitors in the evolution of the proposed BH-WD systems. In the example presented in this paper, and in the similar BeWD source studied in our previous paper, the mass of the OB star presents a real challenge in understanding how so much mass transfer would have taken place to evolve these BeWD systems. It is clearly crucial that future observations establish an accurate estimate of their numbers in the family of evolved binary populations.
# Conclusions
We report in this work the discovery, by S-CUBED of a new BeXRB containing a WD compact object, , which underwent a Type-II X-ray outburst over a period of 120 days. Analysis of optical and IR data from Swift, OGLE and SALT confirm the presence of a O9IIIe companion star hosting a CSD, and X-ray data are consistent with a WD at SMC distance. BeWD binary systems are a rare class of object, of which is only the fourth such system detected in the SMC, despite the predicted large population. We suggest that this may be due in part to the super-soft X-ray emission being only detectable using sensitive focused X-ray imaging telescopes, and only during relatively infrequent Type-II outbursts, that typically last of the order of months. This means that the likelihood of discovery of BeWD sources is greatly increased with regular sensitive soft X-ray monitoring, such as that performed by S-CUBED. It is likely that all-sky sensitive soft X-ray monitoring performed by eROSITA will be a significant discovery engine for BeWD binaries, perhaps discovering the first such system in the Milky Way. However, in order to increase the number of confirmed BeWD binaries, continued monitoring of the SMC by surveys like S-CUBED will prove invaluable. | {'timestamp': '2021-09-14T02:13:51', 'yymm': '2109', 'arxiv_id': '2109.05307', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05307'} |
null | null |
null | null |
# Introduction
Successive relaying (SR) is a well known technique in wireless communication, where two half-duplex (HD) relays simultaneously assist the transmission between a source and destination nodes utilizing the full bandwidth, and thus imitating full-duplex (FD) operation. The main idea is that at a given time-instant, the source transmits data to one relay, and simultaneously, the other relay forwards the data that it received from the source in the previous transmission time-instant to the destination. Hence, it becomes quite clear that a main challenge in realizing such a scheme is inter-relay interference (IRI). Thus far, authors have adopted different approaches to deal with IRI, such as channel ordering and rate adaptation, relay interference cancellation and precoding, or by neglecting IRI assuming fixed directional relays or large distance between active relays.
However, the recently proposed reconfigurable intelligent surfaces (RISs) can be a game changer for many wireless applications, including SR as we will demonstrate in this paper. Thanks to their ability in tweaking the wireless environment while maintaining low cost and power requirements, RISs have gained much attention as a strong candidate for future wireless networks. In principle, RISs are similar to FD amplify-and-forward relaying, with the former providing passive beamforming without the need to include active power amplifiers or radio-frequency chains, while the latter has the ability to provide active amplification at the cost of increased hardware complexity and power consumption.
Instead of dealing with conventional active relaying and RISs as two separate technologies, few works have demonstrated the benefits of integrating RISs in active relaying networks. In particular, the works in have shown that RISs can considerably improve the effective rates of conventional relay-based networks. However, in all of these works the authors considered a conventional relay network with a single relay node aided by an RIS, and no SR or IRI was considered in any of these works. In contrast, in this work we investigate the performance of SR networks with the help of two RISs, where we jointly design the passive beamforming weights for the two RISs to provide a spatial (on the air) suppression of the IRI while maximizing the desired signals gains. Our contributions can be summarized as follows:
- We propose a new successive decode-and-forward (DF) relaying network with two RISs, and investigate its effective rate performance. The corresponding max-min optimization problem is formulated and solved via a semidefinite programming (SDP) approach.
- In addition, as SDP schemes are known to suffer from a high computational complexity, we propose an efficient evolutionary scheme based on particle swarm optimization (PSO) to tackle the joint phase-shift design for both RISs. Furthermore, modifications of the PSO are proposed exploiting the structure of the problem to guarantee stability and control the speed and accuracy of convergence.
- Our results demonstrate that RISs can suppress IRI arising from the SR, and a large gain in achievable rates is obtained compared to the cases where only relays or RISs are utilized. Moreover, the proposed PSO scheme is shown to be highly efficient and can achieve a near-optimal performance when compared to the SDP solution.
It is worth highlighting that our proposed PSO scheme is novel in the sense that it was designed specifically to tackle the RIS passive beamforming with guaranteed stability and controlled convergence behavior.
The rest of this paper is organized as follows: The adopted system model is introduced in Section [2](#section2){reference-type="ref" reference="section2"}. Section [3](#section3){reference-type="ref" reference="section3"} deals with the SDP-based beamforming design, while the PSO-based scheme is introduced and explained in detail in Section [4](#section4){reference-type="ref" reference="section4"}. Numerical results are presented and discussed in Section [5](#section5){reference-type="ref" reference="section5"}. Finally, conclusions are drawn in Section [6](#section6){reference-type="ref" reference="section6"}.
*Notations*: Matrices and vectors are denoted by uppercase and lowercase boldface letters, respectively. \((.)^\ast,\ (.)^T,\ \text{and}\ (.)^H\) denote the conjugate, transpose, and Hermitian transpose operators, respectively. In addition, \(\boldsymbol a_i\), \(\left[\boldsymbol A\right]_{:,j}\), and \(\left[\boldsymbol A\right]_{i,j}\) are the \(i\)th row, \(j\)th column, and \(i\)th element of the \(j\)th column of \(\boldsymbol A\), respectively, while \(\left[\boldsymbol x\right]_i\) and \(x_i\) both represent the \(i\)th element of \(\boldsymbol x\) and used interchangeably as appropriate. \(\boldsymbol 0_{N\times M}\) and \(\boldsymbol I_N\) are the \(N\times M\) all-zero and \(N\times N\) identity matrices, respectively. Furthermore, \(\text{diag}\{\boldsymbol a\}\) is a diagonal matrix whose diagonal is the elements of \(\boldsymbol a\), while \(\text{diag}\{\boldsymbol A\}\) is a vector whose elements are the diagonal elements of \(\boldsymbol A\). The trace of \(\boldsymbol A\) is denoted by \(\text{tr}(\boldsymbol A)\), while \(\boldsymbol A \succeq 0\) indicates that \(\boldsymbol A\) is a positive semidefinite matrix. Finally, \(\left|.\right|\) and \(\mathbb E\{.\}\) denote the absolute and expectation operators, respectively.
# System model {#section2}
We consider a time-division duplex scenario where a source node (\(S\)) transmits data to a destination node (\(D\)). Two HD-DF relays \(R_1\) and \(R_2\), located between \(S\) and \(D\), assist the communication by adopting SR to exploit the full bandwidth. Furthermore, two RISs denoted by \(I_1\) and \(I_2\), located near \(R_1\) and \(R_2\), respectively, are utilized to enhance the communication quality and suppress the IRI arising from the SR operation. We assume that each of \(S\), \(D\), \(R_1\) and \(R_2\) are equipped with a single omni-directional antenna, while \(I_1\) and \(I_2\) each has \(M\) reflective elements.
At any given transmission-time slot, one of the two relays operates as a receiver (Rx) to receive the new block of data from \(S\), while the other relay operates as a transmitter (Tx) to forward the decoded message received from \(S\) in the previous time instant to \(D\); see Fig. [\[system\]](#system){reference-type="ref" reference="system"}.[^1] Focusing on the case where \(R_1\) operates as an Rx, and assuming that signals reflected from each RIS more than once have very low powers and thus can be neglected, we can write the received signal at \(R_1\) as: \[\begin{aligned}
y_{r_1} & = \sqrt{p_s} \left(h_{sr_1}+\boldsymbol{h}_{i_1r_1}^T\boldsymbol \Theta_1\boldsymbol{h}_{si_1}+ \boldsymbol{h}_{i_2r_1}^T\boldsymbol \Theta_2\boldsymbol{h}_{si_2}\right)x \nonumber \\ & + \sqrt{p_{r_2}} \left(h_{r_2r_1}+\boldsymbol{h}_{i_1r_1}^T\boldsymbol \Theta_1\boldsymbol{h}_{r_2i_1}+ \boldsymbol{h}_{i_2r_1}^T\boldsymbol \Theta_2\boldsymbol{h}_{r_2i_2}\right)\tilde{x} + w_{r_1}, \end{aligned}\]where \(p_s\) and \(p_{r_2}\) are the transmit powers of \(S\) and \(R_2\), respectively, \(\boldsymbol \Theta_1\in \mathbb C^{M\times M}\) and \(\boldsymbol \Theta_2\in \mathbb C^{M\times M}\) are the diagonal reflection matrices of \(I_1\) and \(I_2\), respectively, while \(h_{sr_1}\in\mathbb C\), \(h_{r_2r_1}\in\mathbb C\), \(\boldsymbol h_{i_1r_1} \in \mathbb C^{M\times 1}\), \(\boldsymbol h_{si_1} \in \mathbb C^{M\times 1}\), \(\boldsymbol h_{i_2r_1} \in \mathbb C^{M\times 1}\), \(\boldsymbol h_{si_2} \in \mathbb C^{M\times 1}\), \(\boldsymbol h_{r_2i_1} \in \mathbb C^{M\times 1}\), and \(\boldsymbol h_{r_2i_2} \in \mathbb C^{M\times 1}\) denote the channels between \(S\rightarrow R_1\), \(R_2\rightarrow R_1\), \(I_1\rightarrow R_1\), \(S\rightarrow I_1\), \(I_2\rightarrow R_1\), \(S\rightarrow I_2\), \(R_2\rightarrow I_1\), and \(R_2\rightarrow I_2\), respectively.[^2] Moreover, \(x\) and \(\tilde{x}\) are the information symbols transmitted from \(S\) and \(R_2\), respectively, with \(\mathbb E\{|x|^2\} = \mathbb E\{|\tilde x|^2\}=1\), and \(w_{r_1} \sim \mathcal {CN} (0, \sigma^2)\) is the additive white Gaussian noise (AWGN) at \(R_1\).[^3]
For more convenience, we express the effective channels between each node (\(S\), \(R_1\), and \(R_2\)) and the two RISs in a single vector. Specifically, we define \(\small \boldsymbol h_{si} = \begin{bmatrix} \boldsymbol h_{si_1}\\ \boldsymbol h_{si_2} \end{bmatrix} \in\mathbb C^{2M\times 1}\), \(\small \boldsymbol h_{ir_1} = \begin{bmatrix} \boldsymbol h_{i_1r_1}\\ \boldsymbol h_{i_2r_1} \end{bmatrix} \in\mathbb C^{2M\times 1}\), \(\small \boldsymbol h_{r_2i} = \begin{bmatrix} \boldsymbol h_{r_2i_1}\\ \boldsymbol h_{r_2i_2} \end{bmatrix} \in\mathbb C^{2M\times 1}\), \(\small \boldsymbol \theta = \begin{bmatrix} \boldsymbol \theta_1 \\ \boldsymbol \theta_2 \end{bmatrix} \in\mathbb C^{2M\times 1}\), where \(\boldsymbol \theta_j = \text{diag}\{\boldsymbol \Theta_j\} \ (j\in\{1,2\})\). Hence, the signal-to-interference plus noise ratio (SINR) at \(R_1\) can be expressed as follows: \[\label{gamma_r} \small \gamma_{r_1} = \frac{p_s \left|h_{sr_1} + \boldsymbol h_{ir_1}^T \boldsymbol \Theta \boldsymbol h_{si}\right|^2}{p_{r_2} \left|h_{r_2r_1}+\boldsymbol h_{ir_1}^T\boldsymbol \Theta \boldsymbol h_{r_2i} \right|^2 + \sigma^2},\]where \(\boldsymbol \Theta = \text{diag}\{\boldsymbol \theta\} \in \mathbb C^{2M\times 2M}\). On the other hand, we can write the received signal at \(D\) (assuming no direct link between \(S\) and \(D\) exists except through the RISs) as follows: \[\small y_d = \sqrt{p_{r_2}} \left( h_{r_2d} + \boldsymbol h_{id}^T \boldsymbol \Theta \boldsymbol h_{r_2i}\right)\tilde{x} + \sqrt{p_{s}} \left( \boldsymbol h_{id}^T \boldsymbol \Theta \boldsymbol h_{si}\right){x} + w_d,\]where \(\small \boldsymbol h_{id} = \begin{bmatrix} \boldsymbol h_{i_1d}\\ \boldsymbol h_{i_2d} \end{bmatrix} \in\mathbb C^{2M\times 1}\) is a vector containing the channel coefficients between \(D\) and both RISs, \(h_{r_2d}\) is the channel coefficient between \(R_2\) and \(D\), and \(w_d \sim \mathcal{CN} (0, \sigma^2)\) is the AWGN at \(D\). Noting that \(x\) represents interference to the destination in this case, we can write the SINR at \(D\) as: \[\label{gamma_d} \small \gamma_d = \frac{p_{r_2} \left|h_{r_2d} + \boldsymbol h_{id}^T \boldsymbol \Theta \boldsymbol h_{r_2i}\right|^2}{p_s \left| \boldsymbol h_{id}^T \boldsymbol \Theta \boldsymbol h_{si}\right|^2 + \sigma^2}.\]In the subsequent time-instant, \(S\) will transmit a new block of data to \(R_2\) which will operate as an Rx, whereas \(R_1\) will forward its received signal from \(S\) in the current time-slot, assuming successful decoding, to \(D\). Note that if the distances between \(\{S,D\}\) and \(\{R_i, I_i\}\) were the same for both \(i\in \{1,2\}\), then the system is symmetric, and thus it is sufficient to consider only a single phase to evaluate the achievable rate (in this case \(R_1\) is the Rx while \(R_2\) is the Tx). Therefore, as we adopt a symmetric system model in this work, we can write the effective rate at \(D\), in bits/s/Hz, as follows: \[\label{rate} \small \mathcal R = \min \left\{\log_2\left(1+\gamma_{r_1}\right), \ \log_2 \left(1+\gamma_d\right)\right\}.\]It is clear from ([\[gamma_r\]](#gamma_r){reference-type="ref" reference="gamma_r"}) and ([\[gamma_d\]](#gamma_d){reference-type="ref" reference="gamma_d"}) that the achievable rate depends on the reflection matrix \(\boldsymbol \Theta\). In the next section, we formulate and solve the joint passive beamforming design problem by utilizing an SDP approach.
# Passive Beamforming Design via Semidefinite Programming {#section3}
In this section, we aim to maximize the achievable rate in ([\[rate\]](#rate){reference-type="ref" reference="rate"}) via an SDP approach. We start by formulating the corresponding optimization problem as follows:
where \(\mathcal M = \{1, 2,..., 2M\}\). Due to the interference terms in ([\[gamma_r\]](#gamma_r){reference-type="ref" reference="gamma_r"}) and ([\[gamma_d\]](#gamma_d){reference-type="ref" reference="gamma_d"}), and the unit-modulus constraint on \(\boldsymbol \Theta\) in ([\[6a\]](#6a){reference-type="ref" reference="6a"}), problem ([\[OP1\]](#OP1){reference-type="ref" reference="OP1"}) is non-convex and cannot be solved in its current form in polynomial time. Therefore, in the following we adopt change of variables and introduce exponential slack variables to overcome the non-convexity of ([\[OP1\]](#OP1){reference-type="ref" reference="OP1"}).
Let \(\small \boldsymbol q_{sr_1} = \begin{bmatrix} \text{diag}\{\boldsymbol h_{ir_1}\} \boldsymbol h_{si}\\ h_{sr_1} \end{bmatrix}\), \(\small \boldsymbol q_{r_2r_1} = \begin{bmatrix} \text{diag}\{\boldsymbol h_{ir_1}\} \boldsymbol h_{r_2i}\\ h_{r_2r_1} \end{bmatrix}\), \(\small \boldsymbol q_{r_2d} = \begin{bmatrix} \text{diag}\{\boldsymbol h_{id}\} \boldsymbol h_{r_2i}\\ h_{r_2d} \end{bmatrix}\), \(\small \boldsymbol q_{sd} = \begin{bmatrix} \text{diag}\{\boldsymbol h_{id}\} \boldsymbol h_{si}\\ 0 \end{bmatrix}\), and \(\small \boldsymbol v = \begin{bmatrix} \boldsymbol \theta \\ 1 \end{bmatrix}\). Moreover, by defining \(\boldsymbol Q_{sr_1} = \boldsymbol q_{sr_1}\boldsymbol q_{sr_1}^H\), \(\boldsymbol Q_{r_2r_1} = \boldsymbol q_{r_2r_1}\boldsymbol q_{r_2r_1}^H\), \(\boldsymbol Q_{r_2d} = \boldsymbol q_{r_2d}\boldsymbol q_{r_2d}^H\), \(\boldsymbol Q_{sd} = \boldsymbol q_{sd}\boldsymbol q_{sd}^H\), and \(\boldsymbol V = \boldsymbol v^\ast \boldsymbol v^T\), we can now equivalently represent \(\gamma_{r_1}\) and \(\gamma_d\) as follows:
However, even with relaxing the rank-one constraint on \(\boldsymbol V\), formulating the max-min optimization problem with the current formulations of \(\gamma_{r_1}\) and \(\gamma_{d}\) would still result in a non-convex objective function. Nonetheless, this can now be dealt with by introducing additional slack variables. In particular, let \(\boldsymbol s = [s_1\ \ s_2]^T\) and \(\boldsymbol u = [u_1\ \ u_2 ]^T\) be exponential slack variables such that
Subsequently, and after dropping the rank-one constraint on \(\boldsymbol V\) via semidefinite relaxation, we can introduce the following optimization problem[^4]
The only non-convex constraints at this stage are ([\[9b\]](#9b){reference-type="ref" reference="9b"}) and ([\[9d\]](#9d){reference-type="ref" reference="9d"}). To overcome this drawback, a first-order Taylor approximation is applied to convert these constraints into convex ones. In particular, we define \(\bar {\boldsymbol u} = [\bar u_1 \ \ \bar u_2]^T\), where \(\bar u_1\) and \(\bar u_2\) are the points around which the linearization is made such that \(e^{u_i} \approx e^{\bar u_i}(u_i-\bar u_i + 1)\). Hence, we can recast problem ([\[OP2\]](#OP2){reference-type="ref" reference="OP2"}) as
It follows that all constraints and objective function at this stage are convex, and problem ([\[OP3\]](#OP3){reference-type="ref" reference="OP3"}) can be solved iteratively using software tools such as CVX. However, we firstly initialize \(\bar u_1\) and \(\bar u_2\) as follows
where \(\tilde{\boldsymbol V}\) is any feasible (i.e. random) solution. Once the algorithm has converged, there is no guarantee that the rank of \(\boldsymbol V^\star\) is equal to \(1\). In such case, the value of the objective function corresponding to \(\boldsymbol V^\star\) represents an upper-bound. Nonetheless, one can apply eigenvalue decomposition (EVD) with Gaussian randomization to approximate a rank-\(1\) near optimal solution. The details are omitted here for brevity, however, it is worth pointing out that such approximation guarantees a minimum accuracy of \(\pi/4\) of the optimal objective value. The steps of the proposed scheme are given in Algorithm [\[Algorithm1\]](#Algorithm1){reference-type="ref" reference="Algorithm1"}.
:::
# Results and Discussion {#section5}
We start by introducing the different narrowband channels and parameters involved in this paper. All channels from and to each RIS are assumed to follow Rician distribution with both line-of-sight (LoS) and non-LoS (NLoS) links, such that \(\boldsymbol h_j = \sqrt{\frac{k_r}{k_r+1}} \boldsymbol h^{\text{los}}_j + \sqrt{\frac{1}{1+k_r}} \boldsymbol h^{\text{nlos}}_j\), where \(k_r\) is the Rician \(K\) factor, \(\boldsymbol h^{\text{los}}_j \in\mathbb C^{M\times 1}\) is the deterministic LoS channel vector for the \(j\)th link, with each element having an absolute value of \(1/d_j^{{\alpha_{\text{los}}}/2}\) and a random phase between \(0\) and \(2\pi\), \(\boldsymbol h^{\text{nlos}}_j \sim \mathcal {CN}(\boldsymbol 0_{M\times 1}, \boldsymbol I_M d_j^{-{\alpha}_{\text{nlos}}})\) is the Rayleigh distributed NLoS channel vector, where \(d_j\) is the distance between the nodes associated with the \(j\)th link, \(j \in \{si_1, i_1r_1, r_2i_1, i_1d, si_2, i_2r_1, r_2i_2, i_2d\}\), while \(\alpha_{\text{los}}\) and \(\alpha_{\text{nlos}}\) represent the path-loss exponents for LoS and NLoS links, respectively. In contrast, \(h_i \sim \mathcal {CN} (0, d_i^{-\alpha_{\text{nlos}}})\) represents the channel coefficient with Rayleigh distribution for links that do not involve any of the RISs, i.e. \(i\in\{sr_1, r_2d\}\). However, for the IRI link, \(h_{r_2r_1}\), we consider Rician channel to demonstrate the efficiency of RISs in suppressing the interference even when a strong LoS channel exists.
The locations of different nodes are set as: \((x_s, y_s) = (0,0)\), \((x_d, y_d) = (100, 0)\), \((x_{r_1}, y_{r_1}) = (50, 25)\), \((x_{r_2}, y_{r_2}) = (50,-25)\), \((x_{i_1}, y_{i_1}) = (50, 30)\), and \((x_{i_2}, y_{i_2}) = (50,-30)\), all in meter units, see Fig. [\[Fig4\]](#Fig4){reference-type="ref" reference="Fig4"}. In addition, we set \(w_1 = w_2 = 2\), \(\sigma^2 = 1\), \(k_{r} = 5\) dB, \(\alpha_{\text{los}} = 2.3\), and \(\alpha_{\text{nlos}} = 3.5\). Equal power allocation was utilized and thus \(p_s = p_{r_2} = \frac{1}{2}p\), and we define the transmit signal-to-noise ratio (SNR) as \(p/\sigma^2\).
Our simulations compare between the following schemes:
- **RISs-assisted SR (upper-bound)**: This refers to the results obtained by SDP without performing EVD with Gaussian randomization, hence, the term upper-bound. Rician fading is assumed for the IRI channel (i.e. \(h_{r_2r_1}\)).
- **SR without RISs**: This case demonstrates the results for SR utilizing only the two relays without the RISs. Rayleigh fading is adopted for all different channels including the IRI link.
- **RISs only**: In this case, we only have two RISs (therefore no SR) and we optimize their phase-shifts to reflect the signal from \(S\) toward \(D\), where the former transmits with a power budget of \(p\) Watts.
- **RISs-assisted SR (PSO)**: This case shows the performance of our proposed PSO scheme, assuming Rician-fading for the IRI link.
In Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}, we depict the achievable rate vs. SNR curves. The RISs-assisted SR notably outperforms both cases of RISs-only scheme and the SR case without RISs. In particular, to achieve an effective rate of \(4\) bits/s/Hz, a gain of \(15\) dB in transmit SNR can be obtained for the proposed RISs-assisted SR over the RISs-only approach, given that the number of reflecting elements per RIS is \(32\). In contrast, the SR case without RISs shows a very poor performance, which highlights the sever impact of the IRI. Moreover, one can also observe from Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"} that there is only a negligible performance gap between the low-complexity PSO scheme and the more complex SDP approach.
In Fig. [\[Fig3\]](#Fig3){reference-type="ref" reference="Fig3"}, we show the convergence behavior of the proposed PSO scheme for different number of reflecting elements and different values of the auxiliary parameter \(\mu\). Regardless of the number of reflecting elements, setting \(\mu\) to a small value (in this case \(\pi/8\)) is preferable as it allows the proposed algorithm to converge toward better solutions that are close to the optimal points. In contrast, larger values of \(\mu\) enable faster exploration of new and more diverse populations, and thus they are preferable when the number of optimization iterations is small. However, they are more likely to fail in finding near optimal solutions due to the rapid velocity changes of candidate particles, especially when the number of reflecting elements is relatively large.
# Conclusion {#section6}
In this work, a new RIS-aided relay network architecture was proposed with SR, where two RISs were deployed near the two HD-DF relays to provide a spatial suppression of IRI while maximizing the gain of desired signals. An upper bound solution for the achievable rate was obtained via the SDP technique without performing Gaussian randomization and EVD. Subsequently, a lower-complexity beamforming design based on PSO with controlled convergence speed was proposed, and a new normalization step was introduced which guaranteed the stability of the designed scheme. Numerical results demonstrated that RISs can provide large improvements on the rate performance of SR networks, even when a strong line-of-sight link for the IRI existed. Finally, the proposed PSO scheme showed its capability in obtaining near optimal solutions. | {'timestamp': '2021-09-14T02:14:33', 'yymm': '2109', 'arxiv_id': '2109.05316', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05316'} |
# Introduction {#s:intro}
Consider a set of \(n\) points in the plane, called *sites*. Each site \(S_i\) supplies a quantity \(q_i\) of a certain unit-priced product that needs to be collected and has an operating time window of \([s_i, e_i]\), where \(0 \leq s_i \leq e_i \leq T\), for some constant \(T > 0\), \(\forall{i = 1, \dots, n}\). The distance \(d_{i,j}\) between sites \(s_i\) and \(s_j\) is the euclidean distance. We are also given a fleet of \(m\) vehicles \(V = \{v_1, \dots, v_m\}\), each having the same capacity \(Q\). All vehicles are assumed to travel at unit speed and have unit fuel consumption per distance travelled. Only one vehicle may visit any given site and all vehicles must start and end their tour at the same given depot \(D(x_D, y_D)\). The goal is to compute, for every vehicle \(v_k\), a route \(r_k\) that collects a quantity \(q^k_i\) of the good such that the total profit of all routes is maximized, where the profit of a route \(r_k\) is the total quantity collected \(\sum_{s_i \in r_k}{q^k_i}\) (called *reward*) minus the total distance travelled via \(r_k\) (called *costs*). We call this problem the Maximum Profit pick-up Routing Problem (MPRP). MPRP was introduced in, which focuses on the case with fixed quantities and only one vehicle per site. In this paper, we consider the case where multiple vehicles may visit one site. We first study the MPRP problem with Multiple Vehicles Per Site and fixed quantities (MPRP-M), in which \(q_i\)'s are constant in time. We also consider the MPRP problem with Multiple Vehicles Per Site and Time-Variable Quantities supplied (MPRP-MVS), in which \(q_i(t)\)'s are 0 for \(t \notin [s_i, e_i]\) and linearly increasing in time for \(t \in [s_i, e_i]\). Compared to the single vehicle versions, this adds the availability constraint that \(\sum_{j, u \leq t}{q_{iju}} \leq q_i(t), \forall{i, t}\). However, a vehicle may not visit a site more than once. MPRP is known to be strongly NP-hard. By extensions, all the variants proposed above are also strongly NP-hard. That is, they have no algorithm running in time polynomial in the value of the input unless P = NP. Applications of the results presented in this paper include public transportation in various settings, such as cities, inter-city railway transportation, domestic or international flight, etc. Another important application is in for-profit waste pick-up, in which the reward is proportional to the amount of waste collected.
## Related work {#ss:related-work}
General graph TSP is proven to be hard to approximate within any constant factor. However, TSP on special graphs, such as complete graphs with metric distances ane euclidean distances, do have approximation algorithms. For instance, general metric TSP has a \(O(n^3)\) time \(1.5\)-approximation algorithm by Christofides. Euclidean metric TSP even has a PTAS achieving \(1 + \epsilon\) performance ratio via Arora's \(O(n (\log n)^{O(1/\epsilon)})\) time algorithm. On the other hand, metric TSP was shown to be APX-complete by Papadimitriou et. al and, to date, the best known lower bound is \(123/122\) by Karpinski et. al. Various generalizations of TSP have also been considered. Bansal et. al introduced the Deadline-TSP and the Time Window-TSP problems. In Deadline-TSP, every site \(S_i\) has a deadline \(D_i\), while in Time Window-TSP, every site \(S_i\) has a time window \([s_i, e_i]\). Both of these problems associate every site \(S_i\) with a reward \(q_i\) and ask for a single tour that maximized the total reward collected. For Deadline-TSP, they provide an \(O(\log n)\)-approximation algorithm, while for Time Window-TSP they describe an \(O(\log^2 n)\)-approximation. There are also variants involving multiple vehicles. Fisher introduced the Vehicle Routing Problem with capacity constraint (VRP), in which a fleet of \(m\) vehicles is given, and each customer \(S_i\) has a demand \(q_i\) of a certain product. Their solution uses an iterative lagrangian relaxation of the constraints. Later, Fisher et. al introduced the Vehicle Routing Problem with Time Windows and capacity constraints (VRPTW), in which customers have a demand \(q_i\) of the product, as well as operating time windows, and the \(m\) vehicles have non-uniform capacity constraints. They solve this problem using a linear programming approach. Golden et. al studied the Capacitated Arc Routing Problem (CARP) with uniform vehicle capacity where edges, rather than vertices, have customer demands, and the goal is to minimize the travelled distance subject to meeting all demands. They prove that CARP can be approximated within a constant factor when the triangle inequality is satisfied. Later, van Bevern extended the result to general undirected graphs. Wang et. al considered the Multi-Depot Vehicle Routing Problem (MDVRPTW) with Timw Windows and Multi-type Vehicle Number Limits. The main difference to our problem is that the goal is to minimize the number of vehicles used (if feasible) or maximize the number of visited customers (if infeasible). Although they solve the problem using a genetic algorithm approach, their algorithm is iterative and has no performance guarantee with respect to the approximation factor. Versions of vehicle routing with multiple vehicles per customer have also been considered. In, Drexl does a comprehensive survey of variations of vehicle routing with multiple synchronization constraints (VRPMS), e.g. in which a customer may be visited by two vehicles (VRPTT), or by 3 or more vehicles. Armaselu and Daescu solved the fixed-supply, single vehicle version of MPRP. They provide two APXs for MPRP, both running in \(O(n^{11})\) time, assuming \(q_i(l_i) \leq \alpha q_j(l_j), \forall{i, j = 1, \dots, n}\), for some constant \(\alpha > 1\). The best performance ratio, \(15 \log T\), among the two algorithms, is achieved by an approach involving well-separated pair decompositions.
## Our contributions {#ss:structure}
We give a constant-factor approximation algorithm for MPRP-M and MPRP-MVS, which uses clever reductions to some known problems. Specifically, we solve MPRP-M in \(O(n^{11})\) time for an approximation ratio of \(\simeq 44 \log T\) and MPRP-MVS in \(O((\frac{n}{\epsilon})^{11})\) time within an approximation ratio of \(44 \log T (1 + \epsilon)(1 + \frac{1}{1 + \sqrt{m}})^2\). The rest of the paper is structured as follows. In Section [2](#s:mprp-m){reference-type="ref" reference="s:mprp-m"}, we describe our MPRP-M algorthm, and then, in Section [3](#s:mprp-mvs){reference-type="ref" reference="s:mprp-mvs"}, we describe our solution to MPRP-MVS. Finally, in Section [4](#s:conclusion){reference-type="ref" reference="s:conclusion"}, we conclude and list some future directions.
# MPRP-M {#s:mprp-m}
We come up with a reduction from MPRP-M to MPRP. Let \(d(0, i) = d(i, 0)\) be the distance from the depot to site \(S_i, \forall{i: 1 \leq i \leq n}\). An instance \(I\) of MPRP (or MPRP-M) can be described as a set \(I = (S = \{S_i: 1 \leq i \leq n\}, D = (x_0, y_0), d = \{d(i, j): 0 \leq i < j \leq n\}, T, Q, m)\), where \(S_i = (x_i, y_i, s_i, e_i, q_i)\). A solution \(Sol\) to \(I\) is a set \(Sol = (R, P)\). Here \(R\) is a routing function assigning, to every pair of sites or depot \((i, j), 0 \leq i, j \leq n\), a number \(k(i, j): 0 \leq k(i, j) \leq m\), denoting the index of the vehicle operating the path from \(S_i\) to \(S_j\), or 0 if no such vehicle exists. \(P\) is a pickup function assigning, to every site \(S_i\) on vehicle \(v_k\)'s route, a quantity \(q_{k,i} \leq q_i, Q\) to be picked up by \(v_k\). A solution \(Sol'\) to \(I'\) is a routing function with the same properties as the function \(R\) in\(Sol\). Denote by \(MPRP\) (resp., \(MPRP-M\)) the set of MPRP (resp., MPRP-M) instances. Starting from \(I = (\{(x_i, y_i, s_i, e_i, q_i): 1 \leq i \leq n\}, (x_0, y_0), \{d_{i,j}: 0 \leq i < j \leq n\}, T, Q, m) \in MPRP-M\), let \(I' = (\{(x_i, y_i, s_i, e_i, q_i): 1 \leq i \leq n\}, (x_0, y_0), \{d_{i,j}: 0 \leq i < j \leq n\}, T, Q, m) \in MPRP\). That is, \(I'\) has *the same nodes and distances* as \(I\). We solve \(I'\) using the algorithm in and denote by \(Sol'\) the solution. We transform the solution \(Sol'\) to an MPRP instance \(I'\), into a solution \(Sol\) to an MPRP-M instance \(I\), as follows. For every vehicle \(v_k\) of \(I'\), let \(Q_k\) be the total quantity collected by \(v_k\) from its assigned tour \(\tau'_k \in Sol'\). It may be possible that some sites in \(\tau'_k\) need to be visited multiple times in \(I\), producing a new tour \(\tau_k \in Sol\). To figure out which ones need to, we look at tuples of sites \((S_u, S_{u'}, S_i, S_{i'}, S_j, S_{j'}) \in S^2 \times (S \cup D)^4\) in \(I'\) where \[\begin{cases} S_u \in \tau'_k\\ S_{u'} \in \tau'_{k'}\\ S_i, S_j\( consecutive in \)\tau'_k\\ S_{i'}, S_{j'}\( consecutive in \)\tau'_{k'} \end{cases}\] The goal is to maximize the total amount picked up from \(\tau'_k\) and \(\tau'_{k'} \cup \{S_u\}\). Denote by \(q_{u', k'}\) the quanity assigned to be picked up by \(v_{k'}\) from \(S_{u'}\), by \(q_{u', k}\) the quanity assigned to be picked up by \(v_{k}\) from \(S_{u'}\), and by \(q_{u, k}\) the quanity assigned to be picked up by \(v_{k}\) from \(S_{u}\). We have \[\begin{cases} Q_k + q_{u, k} + q_{u', k}-q_{u'} \leq Q\\ Q_{k'} + q_{u', k'} \leq Q\\ q_{u', k} + q_{u', k'} \leq q_{u'}\\ q_{u, k} \leq q_u \end{cases}\] and our goal is to maximize \(q_{u', k} + q_{u', k'} + q_{u, k}\). To do that, we set \(x = q_{u, k}, y = q_{u', k}, z = q_{u', k'}\), and we come to the following linear program.
\[\begin{cases} x \leq q_u\\ y + z \leq q_{u'}\\ z \leq Q-Q_{k'} \end{cases}\] After solving the linear program, we check the additional constraint \(x + y \leq Q-Q_k + q_{u'}\). If it is satisfied, we assign \(v_{k'}\) to pickup \(z^*\) from \(S_{u'}\), and \(v_k\) to pickup \(x^*\) from \(S_u\) and \(y^*\) from \(S_{u'}\) in \(Sol\), where \((x^*, y^*, z^*)\) is a solution to the linear program. In order for \(Sol\) to achieve a better profit than \(Sol'\) through this re-assignment, the following need to hold. Let \(S_u < S_v\) denote that \(S_u\) is visited before \(S_v\) in \(\tau'_{k'}\). \[\begin{cases} x^* + y^* + z^*-q_{u'} > d(i', i) + d(i, i'')-d(i', i'')\\ d(i', i) + d(i, i'')-d(i', i'') \leq \min\{t_{k', v'}-s_{v'}: S_{v'} < S_{i'}\} + \min\{e_{w'}-t_{k', w'}: S_{j'} < S_{w'}\} \end{cases}\] where \(t_{k, i}\) is the time when \(v_k\) visits \(S_i\). That is, insertion of \(S_i\) into \(\tau_{k'}\) in \(Sol\) does not introduce time window violations See Figure 2 for an illustration of this re-assignment. For each vehicle \(v_k\) and site \(S_{u'} \in \tau'_k\), we look at all possible tuples \((S_u, S_i, S_{i'}, S_j, S_{j'}) \in S \times (S \cup D)^4\) satisfying the abovementioned additional constraints and solve the linear program to find the optimal quantities to pickup, then select the tuples that maximize the profit gain after re-assignment, and then perform the re-assignment. One arising concern is that the order in which the pairs \((v_k, S_{u'})\) are selected produces different results. However, while this may happen, it may not impact the optmiality of the re-assignment by more than a constant factor, as we shall see below.
Thus, we first run a MPRP solver to obtain a routing \(\tau'\) and then select the pairs \((v_k, S_{u'})\) in the order given by each \(\tau'_k\) for every \(k\). For each of the \(O(m n)\) pairs, we inspect all possible \(O(n^5)\) tuples in \(O(n)\) time per tuple, which is the time required to verify constraint 2. That is, our reduction is done in \(O(m n^7)\) time. As for the correctness of the reduction, suppose there exists a routing \(\tau''\) for \(I'\) yielding a profit increased by more than a factor of 4 compared to \(\tau'\). In order to do that, \(\tau''\) must pickup a quantity \(q''_{u'} >4 q_{u'}\) from some \(S_{u'}\). However, this implies that either \(S_{u'}\) is not re-assigned in \(\tau\) and thus gives an increased profit for \(I'\), or is re-assigned to one of \(S_u, S_i, S_{i'}, S_j, S_{j'}\) and thus gives an increased profit for one of these sites. Both options lead to a contradiction. Now suppose there exists a routing \(\tau''\) for \(I\) yielding an increased profit compared to \(\tau\). In order for that to happen, either some site \(S_{u'}\) that was never added or involved in a re-assignment picks up a quantity \(q''_{u', k} > q_{u', k}\), implying an increased profit for \(q''_{u'} > 4 q'_{u'}\) for \(S_{u'}\) in \(\tau'\), or some site \(S_{u'}\) involved in a re-assignment with \((S_u, S_i, S_{i'}, S_j, S_{j'})\) picks up a quantity \(q''_{u', k} > 4 q_{u', k}\), implying a quantity was picked up from some site (wlog assumed to be \(S_u\)) in \(\tau''\) which is more than 4 times the one picked from the samee site in \(\tau'\). In both cases, a contradiction follows due to Lemma 1. Hence, we have reduced MPRP-M to MPRP within an approximation ratio of 4 in \(O(m n^7)\) time. By using the \(\simeq 11 \log T\)-approximation for MPRP in for \(O(n^{11})\) time, we thus solve MPRP-M in \(O(n^{11})\) time. We have proved the following result.
# MPRP-MVS {#s:mprp-mvs}
In MPRP-MVS, quantities \(q_i\) supplied at sites vary linearly as a function of time, i.e. \(q_i(t) = q_i(e_i) \frac{t-s_i}{e_i-s_i}\). We adapt the algorithm in for solving MPRP-VS, to work in our case, by using a similar reduction as the one from MPRP-M to MPRP described in the previous section, to reduce MPRP-MVS to MPRP-VS. Given an instance \(I = (S = \{S_i: 1 \leq i \leq n\}, D = (x_0, y_0), d = \{d(i, j): 0 \leq i < j \leq n\}, T, Q, m)\) of MPRP-MVS, we apply the reduction in to obtain an instance \(I'' \in MPRP-M\), and then solve MPRP-M using the algorithm in Section [2](#s:mprp-m){reference-type="ref" reference="s:mprp-m"}. Again, this is not straightforward since the approach in depends on single vehicle assignment which is not the case here. Thus, we need a more insightful construction of \(I''\) than directly applying the algorithm in. For each site \(S_i\), we split \([s_i, e_i]\) into \(N = 1 + \frac{\ln(1/\alpha)}{\epsilon}\) intervals, for some \(\epsilon > 0\), where \(\alpha > 1\) is the smallest constant such that \(q_i(e_i) \leq \alpha q_j(e_j), \forall{i, j}\). Denote these intervals by \(L_{i, l}: 1 \leq l \leq N\), i.e. \(L_{i,l} = [s_i + \frac{e_i-s_i}{(1+\epsilon)^{N-l+1}}, s_i + \frac{e_i-s_i}{(1+\epsilon)^{N-l}}]\). After performing this split, we construct a set of sites \(S'\) where, for each original site \(S_i\), \(S'\) contains \(N\) sites \(S''_{i, \tau}, 1 \leq \tau \leq N\). To each newly constructed site \(S''_{i, l}\), we assign a quantity \(q'_{i, l} = q_i(e_i) \frac{l-0.5}{N-1}\) and a time window \(I_{i,l}\). We then run the MPRP-M Option 1 algorithm described in Section [2](#s:mprp-m){reference-type="ref" reference="s:mprp-m"} on the transformed instance \(I'' = (S'', D, d, T, Q, m)\) (note that \(I'' \in MPRP-M\) since the sites in \(S''\) now supply constant quantities). Denote by \(\tau''\) the routing obtained by running the MPRP-M algorithm on \(I''\). We now analyze the properties of \(\tau''\) in order to put a bound on its performance ratio. We know from that an optimal MPRP algorithm run on \(S'\) may collect a quantity at least \(\frac{1}{1 + \epsilon}\) as much as an optimal MPRP-VS algorithm run on \(S\). Since for \(m = 1\), an instance of MPRP-M also belongs to MPRP, and an instance of MPRP-MVS is also an instance of MPRP-VS, we get the following.
By running the \(\frac{1}{44 \log T}\)-optimal MPRP-M algorithm in Section [2](#s:mprp-m){reference-type="ref" reference="s:mprp-m"} on \(I''\), we get the following a profit at least \(\frac{1}{44 \log T (1 + \frac{1}{1 + \sqrt{m}})}\) as high as an optimal MPRP-MVS algorithm when run on \(I\). Since travel costs are less than the rewards generated by the quantities collected, a performance ratio of \(\frac{1}{1 + \epsilon}\) in the quantities induces a performance ratio of \(\frac{1}{1 + \epsilon}\) in the profits as well. Putting this together with the lemma above, we get the following.
Note that the reduction described before Lemma 2 takes \(O(\frac{n}{\epsilon})\) time. Then, running MPRP-M on \(I''\) which has \(|S''| = O(\frac{n}{\epsilon})\) sites, takes \(O((\frac{n}{\epsilon})^{11})\) time. Thus, we have proved the following result.
# Conclusions and Future Work {#s:conclusion}
We solve the Multiple Vehicles per Site versions of the Maximum-Profit Routing Problem, specifically, the Fixed-Supply version and the Time-Variable Supply version. We leave for future consideration probablisitc approaches for all versions of MPRP, e.g. algorithmic solutions whose output is, with high probability, within a certain bound of the optimum for the given instance. Proving negative results or lower bounds, e.g. inaproximability within a certain ratio, would also be of interest. | {'timestamp': '2021-09-14T02:18:16', 'yymm': '2109', 'arxiv_id': '2109.05404', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05404'} |
null | null |
# Introduction {#sec:intro}
Many researches show that incorporating role-specific language style makes a dialogue system more human-like and attractive . As shown in Figure [\[fig:application\]](#fig:application){reference-type="ref" reference="fig:application"}, in stylistic dialogue generation, the responses are expected to be not only relevant to the context, but also consistent with the desired style. Recently, on the basis of parallel data of an utterance and its stylized version, some studies show the effect of stylistic dialogue generation to account for emotions , response attitudes , personas, and gender identification. . Despite their success, previous studies are generally conducted in a fully supervised setting requiring stylistic dialogue corpus. However, it is expensive and not always feasible to create such a corpus, due to the lack of parallel data which consists of many triplets in {style, context, response} format. For a target style, previous work either looks for specific data sources such as movie transcripts , or recruits human crowd-workers to create corpora . On the other hand, although most state-of-the-art business chatbots are built upon information retrieval technology, there isn't any practical way to train a stylistic retrieval-based dialogue agent. The only way to prevent the dialogue agent from contradicting its desired style is to carefully annotate all the corpus and deliberately filter out those contradicted pairs, which is expensive and inefficient. In this paper, we present a light-weight approach that enables a *generic* retrieval-based dialogue system (which does not embed any persona) to mimic the language style of a specific role. Our framework only requires minimum human inputs (e.g., 500 keywords) for each style, and does not need any paralleled data. Instead of building an entirely new dialogue system for each style from scratch, our framework functions as a "patch" to stylize a generic retrieval-based dialogue system. The main idea is to rewrite a generic dialogue repository into a stylistic one. Specifically, we consider that a target style is represented by a set of jargon phrases (words that are frequently used by a particular group), and propose a jargon-based rewriting schema to tint the responses in dialogue repository with the desired style. One problem in this rewriting schema is that the jargon in the rewritten stylistic responses may have never occurred in the original repository, and the re-ranking model trained on a generic repository may not understand them. To alleviate this problem, we further propose an alignment process to align each stylized response to an internal response that substitute the jargon phrases with synonyms in plain language. In the re-ranking process the model computes matching degrees based on these internal responses and then the aligned stylized responses are proposed as the final output. To evaluate our framework, we implement five dialogue agents with distinct personas and styles--trendy teenagers who prefer informal expression and slang, computer junkies who adopt many technical terms in their language, physics fans who frequently use physics jargon, Pokémon fans who like using game jargon, and martial arts fans who frequently mention tricks or characters in martial arts novels. The stylistic dialogues are evaluated by both human annotators and automatic metrics. Extensive experiments show that our system properly balances the trade-off between relevance and style intensity, and yields responses with better comprehensive performance. Online A/B test indicates that our implemented chatbot with the trendy teenager character setting improves users engagement. This paper has three main contributions:
1. We endow the capability of responding in specific language style to retrieval-based dialogue systems. More importantly, this method can be used to patch any generic dialogue systems for various personas, without any extra fine-tuning or re-training of dialogue models.
2. Our method gets rid of the dependency of paralleled data. Instead we only need a small set of jargon phrases and synonyms to rewrite the dialogue repository.
3. Offline experimental results and online A/B test show the superiority of our approach, as well as the effect of stylistic response on user engagement.
In the remaining part, we review related work. Then we describe our approach in details. Experimental results of both offline evaluation and online A/B test are presented finally.
# Related Work {#sec:related-work}
## Retrieval-based Dialogue Systems
Recently there has been growing interest in building intelligent chatbots to complete tasks or chitchat. This paper focuses on retrieval-based dialogue system that first recall several response candidates from dialogue corpus via lexical-similarity between context, and then learns a semantic matching function between the conversation context and response candidates for response selection Such systems generally produce more fluent and diverse responses and are widely used in commercial chatbots .
## Persona-based Dialogue System
introduce persona-based dialogue models which capture speaker's role and represent it as distributed embeddings. Approaches have been proposed to learn a persona model from user profiles and personalized dialogue . However, most persona-based models focus on the consistency of dialogue with the persona, but do not address the language style problem, which is an important aspect of persona-based dialogue.. Some recent studies focus on specific language styles. introduce a Star Trek chatbot that imitates the speaking style of characters in the show. The chatbot is composed of two models that handle Star Trek style input and everyday conversation separately. propose a set of models for generating polite dialogues. The models are guided by a politeness classifier in generating responses. first extract a non-stylistic prototype from a generic dialogue system and then generate stylistic response via GPT2. and attempt to using monolingual stylistic data to increase the style intensity of dialogue response. We notice that all these studies focus on a single style, and are difficult to extend to other styles. Compared to these methods, our framework is retrieval-based, and can be used cover for a broader range of personas due to its light-weight nature.
## Text Style Transfer
Style transfer of text is a related task of stylizing dialogue systems. The task is transferring a sentence to a specified language style, while keeping the content of the sentence unchanged. A representative approach is learning disentangled latent representations of content and style. propose a cross-aligned auto-encoder with shared content and separated style representations. adversarially combine variational auto-encoder and style classifier. Several studies leverage adversarial networks with different discriminators and training objectives . Another approach separates content and style via explicit sequence editing. show that style transfer can be largely achieved by lexical substitution. propose a "delete, retrieve, and generate" framework that replaces style-related words and re-generate the sentence for better fluency. Similar approaches based on pre-trained language model or reinforcement learning are also explored . Our approach to stylize dialogue is consistent with the above studies (e.g. with the observation of on lexical substitution). However, we do not employ a heavy mechanism to implement the transfer but exploit a small set of style-related terms to generate stylized sentences. We compare with and in experiments.
# Our Proposed Framework
## Main Idea
Our proposed framework is shown in Figure [\[fig:framework\]](#fig:framework){reference-type="ref" reference="fig:framework"}. To construct a stylized retrieval-based dialogue system, our main idea is to rewrite the source corpus \(\mathcal{X}\) to a stylized corpus \(X^Y\). The rewriting is based on a jargon-synonym set that contains about 500 jargon phrases and their synonyms in plain language (annotated by human). The first step is **Response Rewriting** that transforms every response \(r\) in \(\mathcal{X}\) to a stylized response \(r^Y\) with jargon phrases via a style transfer model. After response rewriting, two main challenges still remain: 1) The ranking model trained on a generic repository may not understand the stylized responses \(r^Y\), and 2) the stylized responses \(r^Y\) may be irrelevant to the original context \(c\). To alleviate these problems, we further propose two modules, **Response Alignment** module and **Context Rewriting** module, respectively. The former attempts to align \(r^Y\) to a synonymous response \(r'\) in plain language through the jargon-synonym set, so that the reranking are based on \(r'\) instead of the \(r^Y\). For the latter, we rewrite the context of \(c\) to \(c'\) by adding some context words that are related to the jargon so that \(c'\) is more relevant to the stylized response. For example, for the jargon "Markov chain", we rewrite its origin context *"Isn't life hard for everyone?"* to *"Isn't life hard for everyone? predict; history; trend; uncertain; short-term"*. After Rewriting and Alignment, we can "patch" a generic retrieval-based dialogue system to adapt new language styles. As shown in Figure [\[fig:framework\]](#fig:framework){reference-type="ref" reference="fig:framework"}, after replacing the all context-response pairs \((c, r)\) in a generic conversation repository \(\mathcal{X}\) to stylized pairs \((c', r')\), the dialogue system searches for stylized contexts \(c'\) and selects a synonymous response \(r_{i}^{'}\) as an intermediate output. Then, the corresponded stylized response \(r_i^Y\) is used as the response of input conversation.
## Jargon phrases & synonyms {#sec:Jargon phrases}
We collect a set of jargon phrases from topic-specific websites and ask human experts to annotate them with synonyms in plain language. For instance, for the trendy teenager persona, we collected a list of Internet slang words from Fanjian[^1], a website maintaining a large database of Chinese Internet slang. We took a slang word as a jargon phrase \(s^Y\). For each slang word, three human annotators were asked to assign a synonymous phrase \(s^X\) in plain language. The annotators reviewed and discussed each other's annotation until they reached an agreement. Insulting slang words were removed. Samples of the phrases are presented in Table [4](#tab:sample-align-phrase){reference-type="ref" reference="tab:sample-align-phrase"}. The synonym phrases mostly explain certain aspects of the jargon phrases in plain language. We collected on average 523 parallel jargon phrase-synonym pairs per style.
# Approach
In this section, we introduce detailed procedures of **Response Rewriting**, **Response Alignment**, and **Context Rewriting**, given a set of jargon phrases and synonyms. The specified language style is implemented in the first step. The second step guarantees the synonymity of \(r'\)and \(r^Y\). The third one fixes the relevance between \(c'\) and \(r'\). We further describe rewriting response and rewriting context in details.
## Response Rewriting {#sec:main-corpus-response}
In this section, we introduce detailed procedures of rewriting responses by jargon phrases. Given an original response \(r\), the goal of rewriting is to obtain a fluent and stylized response \(r^Y\). A trivial way to achieve the goal is letting \(r^Y\) be a common sentence like "I am multi-threaded" for any \(r\). To avoid this, we require that \(r^Y\) has as many overlapping words with \(r\) as possible so that it is more specific. We generate a number of candidates for \(r^Y\), and then filter out ungrammatical ones by a fine-tuned BERT model.
### Generating Candidate Response {#sec:main-corpus-response-candidate}
Given a response \(r = \{w_1, w_2,..., w_n\}\), which is a sequence of words, we select a sub-sequence \(s_{i,j} = \{w_i, w_{i+1},..., w_j\}\) and substitute it with a jargon phrase \(s^Y\). The resulting response \(r^Y = \{w_1,..., w_{i-1}, s^Y, w_{j+1},..., w_n\}\) is proposed as a candidate for further filtering. This process is similar to that used in existing style transfer models . These models substitute style-related words and keep others, so that the content is unchanged. To increase the coverage, we allow arbitrary words to be substituted. This relaxation yields \(\mathcal{O}(n^2)\) possible substitutions, so we adopt the following constraints for pruning:
- #### Word constraint:
We only consider words and numbers for substitution. Punctuation marks are not substituted, because they play an important role in grammar. This constraint avoids generating grammatical errors by the change of punctuation.
- #### Length constraint:
We try to keep \(s_{i,j}\) short, so that most words of \(r\) are preserved. This choice is made in order to avoid altering too much the initial sentence. Specifically, we limit the length of \(s_{i,j}\) to be less than 5 words.
- #### Context constraint:
We also require the sub-sequence \(s_{i,j}\) to share the same preceding or succeeding words with a jargon phrase. For example, suppose we find sentences like "\... is fully multi-threaded\..." in the stylized corpus \(\mathcal{T}\). We may consider applying the substitution only if \(s_{i,j}\) succeeds the words "is fully". This constraint helps filter out most of obviously ungrammatical candidates.
### Response Filtering {#sec:main-corpus-response-filter}
We further filter stylized response candidates that are not fluent via a language model. Since a stylized response candidate is not an entirely new sentence, we only need to estimate the fluency of the substituted part. We start with the pre-trained BERT model , and fine-tune it with a non-conversational stylized corpus \(\mathcal{T}\). We mask the jargon phrases in \(\mathcal{T}\) and train the model to recover them. Then we use this model to predict the probability of \(s^Y\) in a given \(r^Y\): \(p(s^Y|r^Y)\). The candidate responses with low \(p(s^Y|r^Y)\) will be filtered.
## Aligning Responses
After the **Response Rewriting** process, it is possible that the generated stylized response \(r^Y\) is not synonymous to the original response \(r\). Moreover, the re-ranking model in the generic dialogue system may not understand \(r^Y\) as well. To alleviate these problems, we substitute stylized phrases in \(r^Y\) with their synonyms in plain language. That is, for each \(s^Y\) in \(r^Y\), we substitute \(s^Y\) with the synonymous \(s^X\). The resulting synonymous response \(r'\) is referred to as the modified response. Such direct substitution ensures the synonymity of \(r'\) and \(r^Y\). Note that \(r'\) could be ungrammatical. This is acceptable because \(r'\) is only used internally by the dialogue system to match with user input and is not shown to users.
## Context Rewriting {#sec:main-corpus-context}
Now we have pairs of synonymous response \(r'\) and stylized response \(r^Y\), and a remaining problem is that the \(r'\) and \(r^Y\) may be totally irrelevant to the original context \(c\) after style transfer. So we attempt to modify \(c\) to a more relevant one, \(c'\), regarded as the stylized context. To obtain \(c'\), we select keywords similar to the stylized phrase \(s^Y\) to append to \(c\). For example, given \(s^y\)="being multi-threaded", the stylized context \(c'\) could be \[\(c\) + "simultaneously" + "moreover"\]. We considered several choices of the keywords, including frequently co-occurred words, nearest words in word embedding space, and simply \(s^X\) itself. We found fastText embedding yields more reasonable keywords. We choose the top 5 words with the highest cosine similarity to the stylized phrase \(s^Y\).
# Evaluation {#sec:eval}
We conducted both offline experiment and online A/B test to evaluate the performance of our framework. Each system is given an utterance in the testing set as input. The output response is evaluated in terms of its quality.
## Dataset[^2] {#Sec:eval-dataset}
To train a stylized dialogue system, our framework does not need paralleled data. Instead, we need a generic dialogue corpus \(\mathcal{X}\), a stylized non-conversational corpus \(\mathcal{T}\), and stylized jargon phrase-synonymous pairs.
#### Source dialogue corpus:
We use the dialogue corpus collected by from Sina Weibo, a popular twitter-like service in China. The contexts are extracted from posts. Comments on the posts are taken as responses. The corpus covers various topics from different users, so we consider it to be in a generic language style. We normalized the tokens and removed non-content words like "retweeted from". Contexts and responses were tokenized by Jieba[^3] toolkit. We sampled 100 posts for testing, and took the remaining 4.4 million context-response pairs as the source dialogue corpus.
#### Stylized non-conversational corpus:
Stylized non-conversational corpus are used for Response Filtering only. For each stylized phrase, we searched Sina Weibo and retrieved the first 500 matched posts. Comments were not retrieved. The posts are pre-processed similarly to the source dialogue corpus. We collected five corpora for the Internet slang style and the computer style.
#### Jargon phrase-synonymous pairs:
As mentioned before, We collected on average 523 parallel jargon phrase-synonym pairs per style.
## Analysis of Style Transfer
In this study, we propose a first study to transfer a generic corpus to a stylized one. To evaluate how effective our transfer approach is, we further compare our style transfer model with two baseline methods, **CAAE** and **DelRetGen**. The evaluation metrics of style transfer are similar to that of stylized dialogue, and we only change the relevance metric to fluency metric. The fluency metric is rated on a scale of 0 to 2 that represent completely non-sense (0), understandable with minor errors (1), and fluent without error (2), respectively. Table [5](#tab:eval-corpus-single){reference-type="ref" reference="tab:eval-corpus-single"} shows the evaluation results. As expected, source responses are rated with perfect fluency and low style degree. They slightly exhibit Internet slang's style, because the dataset is collected from social media. On average, our approach achieves much better fluency and diversity but lower style degree compared to the two baselines. CAAE tends to generate "safe" sentences which make use of only a few target style phrases. To see this, we counted distinct target style phrases in transferred responses. The averaged counts are respectively 81, 172, and 181, for CAAE, DelRetGen, and our approach. DelRetGen improves diversity by involving more target style phrases than CAAE. However, its generation step tends to generate frequent words, limiting its overall diversity. In general, our approach outperforms the baselines on overall performance (FSA, D1, and D2; t-tests with \(p < 10^{-4}\)).
## Online Evaluation
Compared to the generic dialogue system, our framework (as well as other stylized systems) produces more stylized but less relevant responses. We wonder if it has a positive or negative impact on user satisfaction. We conducted A/B test on a commercial social chatbot during one month. We select the trendy teenager style as our testbed. We randomly selected a fraction of users as the treatment group. They were served with the patched system. In case where the patched system fails to find a relevant response, we fallback to the generic system. Users in the control group were always served with the generic system. After the A/B test, we collected 12K conversations for the treatment group. A conversation is a sequence \(\{(u_1, r_1), (u_2, r_2)\}\), where \(u\) is a user utterance and \(r\) is a response. We let \(r_1\) be the patched system's response. The same number of conversations were sampled from the control group. We consider two automatic metrics to measure user satisfaction. The first metric is the sentiment of user's follow-up utterance. Positive sentiment reflects better user satisfaction. We use a sentiment classifier to classify the follow-up utterance \(u_2\). The classifier outputs labels of \(\{-1,0,1\}\), for negative, neutral, and positive sentiment. We take the average label for each group as *Sentiment* in Table [6](#tab:online-eval){reference-type="ref" reference="tab:online-eval"}. The second metric is the fraction of overlapping words between \(r_1\) and \(u_2\) (See *Automatic follow-up* in Table [6](#tab:online-eval){reference-type="ref" reference="tab:online-eval"}). It indicates whether the user is following \(r_1\) or starting a new topic. We also use human annotation for evaluation. We sampled 250 conversations from each group, and asked 5 human annotators to rate user's satisfaction on a scale of-2 to +2:
- **Satisfied (+2)** The user is satisfied with the conversation and shows positive emotion.
- **Follow (+1)** The user understands the response \(r_1\) and follows the topic.
- **Neutral (0)** The user keeps chatting but to ignore \(r_1\).
- **Not-follow (-1)** The user is confused about \(r_1\) or simply exits the chatbot.
- **Unsatisfied (-2)** The user is upset about the conversation and shows negative emotion.
Table [6](#tab:online-eval){reference-type="ref" reference="tab:online-eval"} shows the comparison in the above metrics. The patched system performs consistently better than the baseline system across all metrics.
# Conclusion and Future Work {#sec:conclude}
We propose a light-weight "patching" framework that enables a retrieval-based dialogue system to mimic the language style of a specified character setting. Our framework is flexible and less expensive to apply. Our approach is based on transforming a generic dialogue corpus to a stylized corpus. We design a series of strategies to implement the transformation. Five character settings are implemented. The experiments show that, with several hundreds of manually annotated jargon phrases, our approach is able to produce fluent, relevant, stylized, and diverse responses. A/B test indicates that users are more satisfied about the stylized system. As future work, we plan to improve relevance further and experiment with combining two or more styles in one chatbot. Theoretically our proposed framework is additive, because it provides a way to put different but non-contradictory personalities together. For example, a chatbot could be a computer junkie and a martial arts fan at the same time. When both character settings have response candidates, we can choose one of them by pre-defined priors. It would be interesting to implement such combination and test its effectiveness.
# Introduction {#sec:intro}
Many researches show that incorporating role-specific language style makes a dialogue system more human-like and attractive . As shown in Figure [\[fig:application\]](#fig:application){reference-type="ref" reference="fig:application"}, in stylistic dialogue generation, the responses are expected to be not only relevant to the context, but also consistent with the desired style. Recently, on the basis of parallel data of an utterance and its stylized version, some studies show the effect of stylistic dialogue generation to account for emotions , response attitudes , personas, and gender identification. . Despite their success, previous studies are generally conducted in a fully supervised setting requiring stylistic dialogue corpus. However, it is expensive and not always feasible to create such a corpus, due to the lack of parallel data which consists of many triplets in {style, context, response} format. For a target style, previous work either looks for specific data sources such as movie transcripts , or recruits human crowd-workers to create corpora . On the other hand, although most state-of-the-art business chatbots are built upon information retrieval technology, there isn't any practical way to train a stylistic retrieval-based dialogue agent. The only way to prevent the dialogue agent from contradicting its desired style is to carefully annotate all the corpus and deliberately filter out those contradicted pairs, which is expensive and inefficient. In this paper, we present a light-weight approach that enables a *generic* retrieval-based dialogue system (which does not embed any persona) to mimic the language style of a specific role. Our framework only requires minimum human inputs (e.g., 500 keywords) for each style, and does not need any paralleled data. Instead of building an entirely new dialogue system for each style from scratch, our framework functions as a "patch" to stylize a generic retrieval-based dialogue system. The main idea is to rewrite a generic dialogue repository into a stylistic one. Specifically, we consider that a target style is represented by a set of jargon phrases (words that are frequently used by a particular group), and propose a jargon-based rewriting schema to tint the responses in dialogue repository with the desired style. One problem in this rewriting schema is that the jargon in the rewritten stylistic responses may have never occurred in the original repository, and the re-ranking model trained on a generic repository may not understand them. To alleviate this problem, we further propose an alignment process to align each stylized response to an internal response that substitute the jargon phrases with synonyms in plain language. In the re-ranking process the model computes matching degrees based on these internal responses and then the aligned stylized responses are proposed as the final output. To evaluate our framework, we implement five dialogue agents with distinct personas and styles--trendy teenagers who prefer informal expression and slang, computer junkies who adopt many technical terms in their language, physics fans who frequently use physics jargon, Pokémon fans who like using game jargon, and martial arts fans who frequently mention tricks or characters in martial arts novels. The stylistic dialogues are evaluated by both human annotators and automatic metrics. Extensive experiments show that our system properly balances the trade-off between relevance and style intensity, and yields responses with better comprehensive performance. Online A/B test indicates that our implemented chatbot with the trendy teenager character setting improves users engagement. This paper has three main contributions:
1. We endow the capability of responding in specific language style to retrieval-based dialogue systems. More importantly, this method can be used to patch any generic dialogue systems for various personas, without any extra fine-tuning or re-training of dialogue models.
2. Our method gets rid of the dependency of paralleled data. Instead we only need a small set of jargon phrases and synonyms to rewrite the dialogue repository.
3. Offline experimental results and online A/B test show the superiority of our approach, as well as the effect of stylistic response on user engagement.
In the remaining part, we review related work. Then we describe our approach in details. Experimental results of both offline evaluation and online A/B test are presented finally.
# Related Work {#sec:related-work}
## Retrieval-based Dialogue Systems
Recently there has been growing interest in building intelligent chatbots to complete tasks or chitchat. This paper focuses on retrieval-based dialogue system that first recall several response candidates from dialogue corpus via lexical-similarity between context, and then learns a semantic matching function between the conversation context and response candidates for response selection Such systems generally produce more fluent and diverse responses and are widely used in commercial chatbots .
## Persona-based Dialogue System
introduce persona-based dialogue models which capture speaker's role and represent it as distributed embeddings. Approaches have been proposed to learn a persona model from user profiles and personalized dialogue . However, most persona-based models focus on the consistency of dialogue with the persona, but do not address the language style problem, which is an important aspect of persona-based dialogue.. Some recent studies focus on specific language styles. introduce a Star Trek chatbot that imitates the speaking style of characters in the show. The chatbot is composed of two models that handle Star Trek style input and everyday conversation separately. propose a set of models for generating polite dialogues. The models are guided by a politeness classifier in generating responses. first extract a non-stylistic prototype from a generic dialogue system and then generate stylistic response via GPT2. and attempt to using monolingual stylistic data to increase the style intensity of dialogue response. We notice that all these studies focus on a single style, and are difficult to extend to other styles. Compared to these methods, our framework is retrieval-based, and can be used cover for a broader range of personas due to its light-weight nature.
## Text Style Transfer
Style transfer of text is a related task of stylizing dialogue systems. The task is transferring a sentence to a specified language style, while keeping the content of the sentence unchanged. A representative approach is learning disentangled latent representations of content and style. propose a cross-aligned auto-encoder with shared content and separated style representations. adversarially combine variational auto-encoder and style classifier. Several studies leverage adversarial networks with different discriminators and training objectives . Another approach separates content and style via explicit sequence editing. show that style transfer can be largely achieved by lexical substitution. propose a "delete, retrieve, and generate" framework that replaces style-related words and re-generate the sentence for better fluency. Similar approaches based on pre-trained language model or reinforcement learning are also explored . Our approach to stylize dialogue is consistent with the above studies (e.g. with the observation of on lexical substitution). However, we do not employ a heavy mechanism to implement the transfer but exploit a small set of style-related terms to generate stylized sentences. We compare with and in experiments.
# Our Proposed Framework
## Main Idea
Our proposed framework is shown in Figure [\[fig:framework\]](#fig:framework){reference-type="ref" reference="fig:framework"}. To construct a stylized retrieval-based dialogue system, our main idea is to rewrite the source corpus \(\mathcal{X}\) to a stylized corpus \(X^Y\). The rewriting is based on a jargon-synonym set that contains about 500 jargon phrases and their synonyms in plain language (annotated by human). The first step is **Response Rewriting** that transforms every response \(r\) in \(\mathcal{X}\) to a stylized response \(r^Y\) with jargon phrases via a style transfer model. After response rewriting, two main challenges still remain: 1) The ranking model trained on a generic repository may not understand the stylized responses \(r^Y\), and 2) the stylized responses \(r^Y\) may be irrelevant to the original context \(c\). To alleviate these problems, we further propose two modules, **Response Alignment** module and **Context Rewriting** module, respectively. The former attempts to align \(r^Y\) to a synonymous response \(r'\) in plain language through the jargon-synonym set, so that the reranking are based on \(r'\) instead of the \(r^Y\). For the latter, we rewrite the context of \(c\) to \(c'\) by adding some context words that are related to the jargon so that \(c'\) is more relevant to the stylized response. For example, for the jargon "Markov chain", we rewrite its origin context *"Isn't life hard for everyone?"* to *"Isn't life hard for everyone? predict; history; trend; uncertain; short-term"*. After Rewriting and Alignment, we can "patch" a generic retrieval-based dialogue system to adapt new language styles. As shown in Figure [\[fig:framework\]](#fig:framework){reference-type="ref" reference="fig:framework"}, after replacing the all context-response pairs \((c, r)\) in a generic conversation repository \(\mathcal{X}\) to stylized pairs \((c', r')\), the dialogue system searches for stylized contexts \(c'\) and selects a synonymous response \(r_{i}^{'}\) as an intermediate output. Then, the corresponded stylized response \(r_i^Y\) is used as the response of input conversation.
## Jargon phrases & synonyms {#sec:Jargon phrases}
We collect a set of jargon phrases from topic-specific websites and ask human experts to annotate them with synonyms in plain language. For instance, for the trendy teenager persona, we collected a list of Internet slang words from Fanjian[^5], a website maintaining a large database of Chinese Internet slang. We took a slang word as a jargon phrase \(s^Y\). For each slang word, three human annotators were asked to assign a synonymous phrase \(s^X\) in plain language. The annotators reviewed and discussed each other's annotation until they reached an agreement. Insulting slang words were removed. Samples of the phrases are presented in Table [4](#tab:sample-align-phrase){reference-type="ref" reference="tab:sample-align-phrase"}. The synonym phrases mostly explain certain aspects of the jargon phrases in plain language. We collected on average 523 parallel jargon phrase-synonym pairs per style.
# Approach
In this section, we introduce detailed procedures of **Response Rewriting**, **Response Alignment**, and **Context Rewriting**, given a set of jargon phrases and synonyms. The specified language style is implemented in the first step. The second step guarantees the synonymity of \(r'\)and \(r^Y\). The third one fixes the relevance between \(c'\) and \(r'\). We further describe rewriting response and rewriting context in details.
## Response Rewriting {#sec:main-corpus-response}
In this section, we introduce detailed procedures of rewriting responses by jargon phrases. Given an original response \(r\), the goal of rewriting is to obtain a fluent and stylized response \(r^Y\). A trivial way to achieve the goal is letting \(r^Y\) be a common sentence like "I am multi-threaded" for any \(r\). To avoid this, we require that \(r^Y\) has as many overlapping words with \(r\) as possible so that it is more specific. We generate a number of candidates for \(r^Y\), and then filter out ungrammatical ones by a fine-tuned BERT model.
### Generating Candidate Response {#sec:main-corpus-response-candidate}
Given a response \(r = \{w_1, w_2,..., w_n\}\), which is a sequence of words, we select a sub-sequence \(s_{i,j} = \{w_i, w_{i+1},..., w_j\}\) and substitute it with a jargon phrase \(s^Y\). The resulting response \(r^Y = \{w_1,..., w_{i-1}, s^Y, w_{j+1},..., w_n\}\) is proposed as a candidate for further filtering. This process is similar to that used in existing style transfer models . These models substitute style-related words and keep others, so that the content is unchanged. To increase the coverage, we allow arbitrary words to be substituted. This relaxation yields \(\mathcal{O}(n^2)\) possible substitutions, so we adopt the following constraints for pruning:
- #### Word constraint:
We only consider words and numbers for substitution. Punctuation marks are not substituted, because they play an important role in grammar. This constraint avoids generating grammatical errors by the change of punctuation.
- #### Length constraint:
We try to keep \(s_{i,j}\) short, so that most words of \(r\) are preserved. This choice is made in order to avoid altering too much the initial sentence. Specifically, we limit the length of \(s_{i,j}\) to be less than 5 words.
- #### Context constraint:
We also require the sub-sequence \(s_{i,j}\) to share the same preceding or succeeding words with a jargon phrase. For example, suppose we find sentences like "\... is fully multi-threaded\..." in the stylized corpus \(\mathcal{T}\). We may consider applying the substitution only if \(s_{i,j}\) succeeds the words "is fully". This constraint helps filter out most of obviously ungrammatical candidates.
### Response Filtering {#sec:main-corpus-response-filter}
We further filter stylized response candidates that are not fluent via a language model. Since a stylized response candidate is not an entirely new sentence, we only need to estimate the fluency of the substituted part. We start with the pre-trained BERT model , and fine-tune it with a non-conversational stylized corpus \(\mathcal{T}\). We mask the jargon phrases in \(\mathcal{T}\) and train the model to recover them. Then we use this model to predict the probability of \(s^Y\) in a given \(r^Y\): \(p(s^Y|r^Y)\). The candidate responses with low \(p(s^Y|r^Y)\) will be filtered.
## Aligning Responses
After the **Response Rewriting** process, it is possible that the generated stylized response \(r^Y\) is not synonymous to the original response \(r\). Moreover, the re-ranking model in the generic dialogue system may not understand \(r^Y\) as well. To alleviate these problems, we substitute stylized phrases in \(r^Y\) with their synonyms in plain language. That is, for each \(s^Y\) in \(r^Y\), we substitute \(s^Y\) with the synonymous \(s^X\). The resulting synonymous response \(r'\) is referred to as the modified response. Such direct substitution ensures the synonymity of \(r'\) and \(r^Y\). Note that \(r'\) could be ungrammatical. This is acceptable because \(r'\) is only used internally by the dialogue system to match with user input and is not shown to users.
## Context Rewriting {#sec:main-corpus-context}
Now we have pairs of synonymous response \(r'\) and stylized response \(r^Y\), and a remaining problem is that the \(r'\) and \(r^Y\) may be totally irrelevant to the original context \(c\) after style transfer. So we attempt to modify \(c\) to a more relevant one, \(c'\), regarded as the stylized context. To obtain \(c'\), we select keywords similar to the stylized phrase \(s^Y\) to append to \(c\). For example, given \(s^y\)="being multi-threaded", the stylized context \(c'\) could be \[\(c\) + "simultaneously" + "moreover"\]. We considered several choices of the keywords, including frequently co-occurred words, nearest words in word embedding space, and simply \(s^X\) itself. We found fastText embedding yields more reasonable keywords. We choose the top 5 words with the highest cosine similarity to the stylized phrase \(s^Y\).
# Evaluation {#sec:eval}
We conducted both offline experiment and online A/B test to evaluate the performance of our framework. Each system is given an utterance in the testing set as input. The output response is evaluated in terms of its quality.
## Dataset[^6] {#Sec:eval-dataset}
To train a stylized dialogue system, our framework does not need paralleled data. Instead, we need a generic dialogue corpus \(\mathcal{X}\), a stylized non-conversational corpus \(\mathcal{T}\), and stylized jargon phrase-synonymous pairs.
#### Source dialogue corpus:
We use the dialogue corpus collected by from Sina Weibo, a popular twitter-like service in China. The contexts are extracted from posts. Comments on the posts are taken as responses. The corpus covers various topics from different users, so we consider it to be in a generic language style. We normalized the tokens and removed non-content words like "retweeted from". Contexts and responses were tokenized by Jieba[^7] toolkit. We sampled 100 posts for testing, and took the remaining 4.4 million context-response pairs as the source dialogue corpus.
#### Stylized non-conversational corpus:
Stylized non-conversational corpus are used for Response Filtering only. For each stylized phrase, we searched Sina Weibo and retrieved the first 500 matched posts. Comments were not retrieved. The posts are pre-processed similarly to the source dialogue corpus. We collected five corpora for the Internet slang style and the computer style.
#### Jargon phrase-synonymous pairs:
As mentioned before, We collected on average 523 parallel jargon phrase-synonym pairs per style.
## Analysis of Style Transfer
In this study, we propose a first study to transfer a generic corpus to a stylized one. To evaluate how effective our transfer approach is, we further compare our style transfer model with two baseline methods, **CAAE** and **DelRetGen**. The evaluation metrics of style transfer are similar to that of stylized dialogue, and we only change the relevance metric to fluency metric. The fluency metric is rated on a scale of 0 to 2 that represent completely non-sense (0), understandable with minor errors (1), and fluent without error (2), respectively. Table [5](#tab:eval-corpus-single){reference-type="ref" reference="tab:eval-corpus-single"} shows the evaluation results. As expected, source responses are rated with perfect fluency and low style degree. They slightly exhibit Internet slang's style, because the dataset is collected from social media. On average, our approach achieves much better fluency and diversity but lower style degree compared to the two baselines. CAAE tends to generate "safe" sentences which make use of only a few target style phrases. To see this, we counted distinct target style phrases in transferred responses. The averaged counts are respectively 81, 172, and 181, for CAAE, DelRetGen, and our approach. DelRetGen improves diversity by involving more target style phrases than CAAE. However, its generation step tends to generate frequent words, limiting its overall diversity. In general, our approach outperforms the baselines on overall performance (FSA, D1, and D2; t-tests with \(p < 10^{-4}\)).
## Online Evaluation
Compared to the generic dialogue system, our framework (as well as other stylized systems) produces more stylized but less relevant responses. We wonder if it has a positive or negative impact on user satisfaction. We conducted A/B test on a commercial social chatbot during one month. We select the trendy teenager style as our testbed. We randomly selected a fraction of users as the treatment group. They were served with the patched system. In case where the patched system fails to find a relevant response, we fallback to the generic system. Users in the control group were always served with the generic system. After the A/B test, we collected 12K conversations for the treatment group. A conversation is a sequence \(\{(u_1, r_1), (u_2, r_2)\}\), where \(u\) is a user utterance and \(r\) is a response. We let \(r_1\) be the patched system's response. The same number of conversations were sampled from the control group. We consider two automatic metrics to measure user satisfaction. The first metric is the sentiment of user's follow-up utterance. Positive sentiment reflects better user satisfaction. We use a sentiment classifier to classify the follow-up utterance \(u_2\). The classifier outputs labels of \(\{-1,0,1\}\), for negative, neutral, and positive sentiment. We take the average label for each group as *Sentiment* in Table [6](#tab:online-eval){reference-type="ref" reference="tab:online-eval"}. The second metric is the fraction of overlapping words between \(r_1\) and \(u_2\) (See *Automatic follow-up* in Table [6](#tab:online-eval){reference-type="ref" reference="tab:online-eval"}). It indicates whether the user is following \(r_1\) or starting a new topic. We also use human annotation for evaluation. We sampled 250 conversations from each group, and asked 5 human annotators to rate user's satisfaction on a scale of-2 to +2:
- **Satisfied (+2)** The user is satisfied with the conversation and shows positive emotion.
- **Follow (+1)** The user understands the response \(r_1\) and follows the topic.
- **Neutral (0)** The user keeps chatting but to ignore \(r_1\).
- **Not-follow (-1)** The user is confused about \(r_1\) or simply exits the chatbot.
- **Unsatisfied (-2)** The user is upset about the conversation and shows negative emotion.
Table [6](#tab:online-eval){reference-type="ref" reference="tab:online-eval"} shows the comparison in the above metrics. The patched system performs consistently better than the baseline system across all metrics.
# Conclusion and Future Work {#sec:conclude}
We propose a light-weight "patching" framework that enables a retrieval-based dialogue system to mimic the language style of a specified character setting. Our framework is flexible and less expensive to apply. Our approach is based on transforming a generic dialogue corpus to a stylized corpus. We design a series of strategies to implement the transformation. Five character settings are implemented. The experiments show that, with several hundreds of manually annotated jargon phrases, our approach is able to produce fluent, relevant, stylized, and diverse responses. A/B test indicates that users are more satisfied about the stylized system. As future work, we plan to improve relevance further and experiment with combining two or more styles in one chatbot. Theoretically our proposed framework is additive, because it provides a way to put different but non-contradictory personalities together. For example, a chatbot could be a computer junkie and a martial arts fan at the same time. When both character settings have response candidates, we can choose one of them by pre-defined priors. It would be interesting to implement such combination and test its effectiveness.
[^1]: http://www.fanjian.net/
[^2]: All datasets will be released.
[^3]: https://github.com/fxsjy/jieba
[^4]: We tried to optimize the original implementation of DelRetGen, but still can not rewrite the corpus in a reasonable time.
[^5]: http://www.fanjian.net/
[^6]: All datasets will be released.
[^7]: https://github.com/fxsjy/jieba
[^8]: We tried to optimize the original implementation of DelRetGen, but still can not rewrite the corpus in a reasonable time. | {'timestamp': '2021-09-14T02:21:09', 'yymm': '2109', 'arxiv_id': '2109.05477', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05477'} |
null | null |
null | null |
null | null |
# Broader Impact {#broader-impact .unnumbered}
The algorithmic processes behind modern search engines, with extensive use of machine learning algorithms, have great power to determine users' access to information . Our research provides evidence that unintentionally using image search models trained either on in-domain image retrieval data sets or massive corpora across the internet may lead to unequal inclusiveness between males and females in image search results, even when the search terms are gender-neutral. This inequity can and do have significant impacts on shaping and exaggerating gender stereotype in people's minds . This work offers new methods for mitigating gender bias in multimodal models, and we regard the algorithms proposed in this paper have the potentials to be deployed in real-world systems. We conjecture that our methods may contribute to driving the development of responsible image search engines with other fairness issues. For instance, we would encourage future works to understand and mitigate the risks arising from other social biases, like racial bias, in image search results. We would also encourage researchers to explore whether the methodology presented in this work could be generalized to quantify and mitigate other bias measures. Our work has limitations. The gender bias measures and the debiasing methods proposed in this study require acquiring the gender labels of images. Our method for identifying the gender attributes of people portrayed in the images is limited: we make use of the contextual cues in the human-annotated captions from the image datasets. The accuracy of such a proxy-based method heavily relies on the coverage of gendered nouns and the inclusiveness of gendered language in the original human annotations. The corruption of gender labels, due to missing gendered words or inappropriate text pre-processing steps, may introduce biases we have not foreseen into the evaluated metrics. Additionally, the gendered word lists are collected from English corpora and may differ in other languages or cultures. It is possible that blind application of our methods by improperly acquiring the gender labels may create image search models that produce even greater inequality, which is very much discouraged. This limitation arises from the unavailability of such sensitive attributes in the source datasets. The lack of relevant data for studying gender bias in image search, and the concerns about how to acquire the gender attributes while preserving the privacy of people concerned, is itself an important question in this area. We believe this research would benefit when richer datasets become available.
# Introduction {#sec:introduction}
Internet information is shaping people's minds. The algorithmic processes behind modern search engines, with extensive use of machine learning, have great power to determine users' access to information . These information systems are biased when results are systematically slanted in unfair discrimination against protected groups . Gender bias is a severe fairness issue in image search. [\[fig:case_study\]](#fig:case_study){reference-type="ref" reference="fig:case_study"} shows an example: given a gender-neutral natural language query "a person is cooking\", only 2 out of 10 images retrieved by an image search model depict females, while equalized exposure for male and female is expected. Such gender-biased search results are harmful to society as they change people's cognition and worsen gender stereotypes . Mitigating gender bias in image search is imperative for social good. In this paper, we formally develop a framework for quantifying gender bias in image search results, where text queries in English[^1] are made gender-neutral, and gender-balanced search images are expected for models to retrieve. To evaluate model fairness, we use the normalized difference between masculine and feminine images in the retrieved results to represent gender bias. We diagnose the gender bias of two primary families of multimodal models for image search: (1) the specialized models that are often trained on in-domain datasets to perform text-image retrieval, and (2) the general-purpose representation models that are pre-trained on massive image and text data available online and can be applied to image search. Our analysis on MS-COCO and Flickr30K datasets reveals that both types of models lead to serious gender bias issues (e.g., nearly 70% of the retrieved images are masculine images). To mitigate gender bias in image search, we propose two novel debiasing solutions for both model families. The specialized in-domain training methods such as SCAN often adopt contrastive learning to enforce image-text matching by maximizing the margin between positive and negative image-text pairs. However, the gender distribution in the training data is typically imbalanced, which results in unfair model training. Thus we introduce a fair sampling (*FairSample*) method to alleviate the gender imbalance during training without modifying the training data. Our second solution aims at debiasing the large, pre-trained multimodal representation models, which effectively learn pre-trained image and text representations to accomplish down-stream applications . We examine whether the representative CLIP model embeds human biases into multimodal representations when they are applied to the task of image search. Furthermore, we propose a novel post-processing feature clipping approach, *clip*, that effectively prunes out features highly correlated with gender based on their mutual information to reduce the gender bias induced by multimodal representations. The *clip* method does not require any training and is compatible with various pre-trained models. We evaluate both debiasing approaches on MS-COCO and Flickr30K and find that, on both benchmarks, the proposed approaches significantly reduce the gender bias exhibited by SCAN and CLIP models when evaluated on the gender-neutral corpora, yielding fairer and more gender-balanced search results. In addition, we evaluate the similarity bias of the CLIP model in realistic image search results for occupations on the internet, and observe that the post-processing methods mitigate the discrepancy between gender groups by a large margin. Our contributions are four-fold: (1) we diagnose a unique gender bias in image search, especially for gender-neutral text queries; (2) we introduce a fair sampling method to mitigate gender bias during model training; (3) we also propose a novel post-processing clip method to debias pre-trained multimodal representation models; (4) we conduct extensive experiments to analyze the prevalent bias in existing models and demonstrate the effectiveness of our debiasing methods.
# Mitigating Gender Bias in Image Search {#sec:methodology}
There are two fashions of multimodal models for the image search task. One is to build a specialized model that could embed image and text into representation vectors with measurable similarity scores. The other is to use general-purpose image-text representations pre-trained on sufficiently big data and compute a particular distance metric. We focus on two representative models, SCAN and CLIP , for both fashions. For the first fashion, we propose an in-processing learning approach to ameliorate the unfairness caused by imbalanced gender distribution in training examples. This approach builds on contrastive learning but extends with a *fair sampling* step. The in-processing solution requires full training on in-domain data examples. For the second fashion, we propose a post-processing feature *clipping* technique to mitigate bias from an information-theoretical perspective. This approach is compatible with pre-trained models and is light to implement without repeating training steps.
## In-processing Debiasing: Fair Sampling {#sec:fairsample}
Image search models in the first fashion are often trained under the contrastive learning framework . For our in-processing debiasing approach, we now explain the two primary components, contrastive learning and fair sampling, within our context.
#### Contrastive Learning
We start by formally introducing the standard contrastive learning framework commonly used in previous works for image-text retrieval. Given a batch of \(N\) image-text pairs \(\batch = \{(v_n,c_n)\}_{n=1}^N\), the model aims to maximize the similarity scores of matched image-text pairs (positive pairs) while minimizing that of mismatched pairs (negative pairs). The representative SCAN model , denoted as \(S(v, c)\) outputting a similarity score between image and text, is optimized with a standard hinge-based triplet loss: \[\begin{aligned}
\mathcal{L}_{i-t} & = \sum_{(v, c) \in \batch} [\gamma-S(v, c) + S(v, \tilde{c})]_+ \\ \mathcal{L}_{t-i} & = \sum_{(v, c) \in \batch} [\gamma-S(v, c) + S(\tilde{v}, c)]_+ \end{aligned}\] where \(\gamma\) is the margin, \(\tilde{v}\) and \(\tilde{c}\) are negative examples, and \([\cdot]_+\) denotes the ramp function. \(\mathcal{L}_{i-t}\) corresponds to image-to-text retrieval, while \(\mathcal{L}_{t-i}\) corresponds to text-to-image retrieval (or image search). Common negative sampling strategy includes selecting all the negatives , selecting hard negatives of highest similarity scores in the mini-batch , and selecting hard negatives from the whole training data . Minimizing the margin-based triplet loss will make positive image-text pairs closer to each other than other negative samples in the joint embedding space.
#### Fair Sampling
One major issue in the contrastive learning framework is that the gender distribution in a batch of image-text pairs is typically imbalanced. Hence, the negative samples will slant towards the majority group, leading to systematic discrimination. To address this problem, we propose a fair sampling strategy. We split the batch of image-text pairs into masculine and feminine pairs based on the image's gender attribute: \[\begin{aligned}
\imageset_\male & = \{v \mid g(v) = \male, (v, c) \in \batch\} \\ \imageset_\female & = \{v \mid g(v) = \female, (v, c) \in \batch\} \\ \imageset_\neutral & = \{v \mid g(v) = \neutral, (v, c) \in \batch\} \end{aligned}\] For every positive image and text pair \((v, c) \in \batch\), we identify the gender information contained in the query \(c\). If the natural language query is gender-neutral, we sample a negative image from the set of male and female images with probability \(\frac{1}{2}\), respectively. Otherwise, we keep the primitive negative sampling selection strategy for keeping the model's generalization on gender-specific queries. Let \(\batch^\ast\) be the batch of gender-neutral image-text pairs, the image search loss with fair sampling is: \[\begin{gathered}
\mathcal{L}_{t-i}^{fair} = \sum_{(v, c) \in \batch^\ast}(\frac{1}{2}\Expectation_{\bar{v} \in \imageset_\male} [\gamma-S(v, c) + S(\bar{v}, c)]_+ \\ \quad + \frac{1}{2}\Expectation_{\bar{v} \in \imageset_\female} [\gamma-S(v, c) + S(\bar{v}, c)]_+) \\ + \sum_{(v, c) \in \batch / \batch^\ast} [\gamma-S(v, c) + S(\tilde{v}, c)]_+ \end{gathered}\] Empirically, we find that if we thoroughly apply the Fair Sampling strategy, the recall performance drops too much. To obtain a better tradeoff, we use a weight \(\alpha\) to combine the objectives \[\alpha \mathcal{L}_{t-i}^{fair} + (1-\alpha)\mathcal{L}_{t-i}\] as the final text-to-image loss function. We do not alter the sentence retrieval loss \(\mathcal{L}_{i-t}\) during training for preserving generalization.
## Post-processing Debiasing: Feature Clipping based on Mutual Information {#sec:prune}
Pre-training methods have shown promising zero-shot performance on extensive NLP and computer vision benchmarks. The recently introduced CLIP model was pre-trained on an enormous amount of image-text pairs found across the internet to connect text and images. CLIP can encode image and text into \(d\)-dimensional embedding vectors, based on which we can use cosine similarity to quantify the similarity of image and text pairs. In this work, we find that the pre-trained CLIP model reaches the state-of-the-art performance but exhibits large gender bias due to training on uncurated image-text pairs collected from the internet. Although released the pre-trained CLIP model, the training process is almost unreproducible due to limitations on computational costs and massive training data. In order to avoid re-training of the CLIP model, we introduce a novel post-processing mechanism to mitigate the representation bias in the CLIP model. We propose to "clip" the dimensions of feature embeddings that are highly correlated with gender information. This idea is motivated by the fact that an unbiased retrieve implies the independence between the covariates (active features) and sensitive attributes (gender) . Clipping the highly correlating covariates will return us a relatively independent and neutral set of training data that does not encode hidden gender bias.
:::
The proposed *clip* algorithm is demonstrated in [\[algo:prune\]](#algo:prune){reference-type="ref" reference="algo:prune"}, and we explain the key steps below. Let \(\Omega = \{1...,d\}\) be the full index set. We use \(V = V_\Omega = [V_1, V_2,..., V_d]\) to represent the variable of \(d\)-dimensional encoding image vectors and \(g(V) \in \{\male, \female, \neutral\}\) to represent the corresponding gender attribute. The goal is to output the index set \(\mathcal{Z}\) of clipped covariates that reduce the dependence between representations \(V_{\Omega/\mathcal{Z}}\) and gender attributes \(g(V)\). We measure the correlation between each dimension \(V_i\) and gender attribute \(g(V)\) by estimating their mutual information \(I(V_i;g(V))\) : \[I(V_I;g(V)) = D_{\text{KL}}(\Probability_{(V_i,g(V))}\|\Probability_{V_i} \otimes \Probability_{g(V)})\] where \(D_{\text{KL}}\) is the KL divergence , \(\Probability_{(V_i,g(V))}\) indicates the joint distribution, \(\Probability_{V_i}\) and \(\Probability_{g(V)}\) indicate their marginals. Next, we greedily clip \(m\) covariates with highest mutual information, and construct \((d-m)\)-dimensional embedding vectors \(V_{\Omega / \mathcal{Z}}\). \(m\) is a hyper-parameter that we will experimentally find to best trade-off accuracy and the reduced gender bias, and we show how the selection of \(m\) affects the performance in [4.3](#sec:trade-off){reference-type="ref" reference="sec:trade-off"}. To project text representations, denoted by variable \(C\), into the same embedding space, we also apply the index set \(\mathcal{Z}\) to obtain clipped text embedding vectors \(C_{\Omega / \mathcal{Z}}\). The clipped image and text representations, denoted by \(\vec{v}^\ast\) and \(\vec{c}^\ast\), will have a relatively low correlation with gender attributes due to the "loss" of mutual information. Then we compute the cosine similarity between image and text by substituting \(\vec{v}^\ast\) and \(\vec{c}^\ast\) into [\[eq:similarity\]](#eq:similarity){reference-type="ref" reference="eq:similarity"}: \[S(v, c) = \frac{\vec{v}^\ast \cdot \vec{c}^\ast}{\|\vec{v}^\ast\| \|\vec{c}^\ast\|}\] Finally, we rank the images based on the cosine similarity between the clipped representations.
# Experimental Setup {#sec:experiment}
## Datasets
We evaluate our approaches on the standard MS-COCO and Flickr30K datasets. Following and, we split MS-COCO captions dataset into 113,287 training images, 5,000 validation images and 5,000 test images.[^2] Each image corresponds to 5 human-annotated captions. We report the results on the test set by averaging over five folds of 1K test images or evaluating the full 5K test images. Flickr30K consists of 31,000 images collected from Flickr.[^3] Following the same split of, we select 1,000 images for validation, 1,000 images for testing, and the rest of the images for training.
#### Identifying Gender Attributes of Images
Sensitive attributes such as gender are often not explicitly annotated in large-scale datasets such as MS-COCO and Flickr30K, but we observe that implicit gender attributes of images can be extracted from their associated human-annotated captions. Therefore, we pre-define a set of masculine words and a set of feminine words.[^4] Following and we use the ground-truth annotated captions to identify the gender attributes of images. An image will be labeled as "male" if at least one of its captions contains masculine words and no captions include feminine words. Similarly, an image will be labeled as "female\" if at least one of its captions contains feminine words and no captions include masculine words. Otherwise, the image will be labeled as "gender-neutral\".
## Models
We compare the fairness performance of the following approaches:
**SCAN** : we use the official implementation for training and evaluation[^5].
**FairSample**: we apply the fair sampling method proposed in [2.1](#sec:fairsample){reference-type="ref" reference="sec:fairsample"} to the SCAN framework and adopt the same hyper-parameters suggested by for training.
**CLIP** : we use the pre-trained CLIP model released by OpenAI.[^6] The model uses a Vision Transformer as the image encoder and a masked self-attention Transformer as the text encoder. The original model produces 500-dimensional image and text vectors.
**CLIP-clip**: we apply the feature pruning algorithm in [2.2](#sec:prune){reference-type="ref" reference="sec:prune"} to the image and text features generated by the CLIP model. We set \(m=100\) and clip the image and text representations into 400-dimensional vectors. Note that SCAN and FairSample are trained and tested on the in-domain MS-COCO and Flickr30K datasets, while the pre-trained CLIP model is directly tested on MS-COCO and Flickr30K test sets without fine-tuning on their training sets (same for CLIP-clip as it simply drops CLIP features).
## Evaluation
#### Gender-Neutral Text Queries
In this study, we focus on equalizing the search results of gender-neutral text queries. In addition to the existing gender-neutral captions in the test sets, we pre-process those gender-specific captions to construct a purely gender-neutral test corpus to guarantee a fair and large-scale evaluation. For every caption, we identify all these gender-specific words and remove or replace them with corresponding gender-neutral words. We show some pre-processing examples in [\[tab:construct-neutral-query\]](#tab:construct-neutral-query){reference-type="ref" reference="tab:construct-neutral-query"}.
#### Metrics
As introduced in [5.2](#sec:measure-gender-bias){reference-type="ref" reference="sec:measure-gender-bias"}, we employ the fairness metric in [\[eq:fairness\]](#eq:fairness){reference-type="ref" reference="eq:fairness"}, Bias\@K, to measure the gender bias among the top-K images. In addition, following standard practice, we measure the retrieval performance by Recall\@K, defined as the fraction of queries for which the correct image is retrieved among the top-K images.
# Debiasing Results
## Main Results on MS-COCO & Flickr30K
We report the results comparing our debiasing methods and the baseline methods in [\[tab:main_result\]](#tab:main_result){reference-type="ref" reference="tab:main_result"}.
#### Model Bias
Although the pre-trained CLIP model is evaluated without fine-tuning, we observe that it achieves a comparable recall performance with the SCAN model on MS-COCO and dominates the Flickr30K dataset. However, both models suffer from severe gender bias. Especially, the \(\bias@10\) of the SCAN model on Flickr30K is \(0.3960\), meaning nearly \(70\%\) of the retrieved gender-specific images portray men and only \(30\%\) portray women. Similarly, the CLIP model achieves \(0.2648\) gender bias on MS-COCO 1K test set, indicating about \(6.4\) out of 10 retrieved images portray men while about \(3.6\) out of 10 portray women. Given that all of the testing text queries are gender-neutral, this result shows that severe implicit gender bias exists in image search models.
#### Debiasing Effectiveness
As shown in [\[tab:main_result\]](#tab:main_result){reference-type="ref" reference="tab:main_result"}, both the in-processing sampling strategy *FairSample* and the post-processing feature pruning algorithm *clip* consistently mitigate the gender bias on test data. For instance, among the top-10 search images, SCAN with FairSample reduces gender bias from \(0.3960\) to \(0.3537\) (decreased by \(10.7\%\)) on Flickr30K. Using the clipped CLIP features for image search (CLIP-clip), the gender bias drops from \(0.2648\) to \(0.2057\) (\(22.3\%\)) on MS-COCO 1K, from \(0.2131\) to \(0.1611\) (\(24.4\%\)) on MS-COCO 5K, and from \(0.3586\) to \(0.2951\) (17.7%) on Flickr30K. For the tradeoff, CLIP-clip sacrifices the recall performance slightly (from \(93.6\%\) Recall\@10 to \(91.3\%\) on Flickr30K). On the other hand, SCAN with FairSample even achieves a comparable recall performance with SCAN.
## Gender Bias at Different Top-K Results
We plot how gender bias varies across different values of \(K\) (1-10) for all the compared methods in [\[fig:gender-bias\]](#fig:gender-bias){reference-type="ref" reference="fig:gender-bias"}. We observe that when \(K < 5\), the gender bias has a higher variance due to the inadequate retrieved images. When \(K \geq 5\), the curves tend to be flat. This result indicates that \(\bias@10\) is more recommended than \(\bias@1\) for measuring gender bias as it is more stable. It is also noticeable that CLIP-clip achieves the best fairness performance in terms of \(\bias@10\) consistently on all three test sets compared to the other models.
## Tradeoff between Recall and Bias {#sec:trade-off}
There is an inherent tradeoff between fairness and accuracy in fair machine learning . To achieve the best recall-bias tradeoff in our methods, we further examine the effect of the controlling hyper-parameters: the weight \(\alpha\) in FairSampling and the number of clipped dimensions \(m\) in CLIP-clip. [\[fig:fairsample_tradeoff\]](#fig:fairsample_tradeoff){reference-type="ref" reference="fig:fairsample_tradeoff"} demonstrates the recall-bias curve with the fair sampling weight \(\alpha \in [0,1]\). Models of higher recall often suffer higher gender bias, but the fairness improvement outweighs the recall performance drop in FairSample models. For example, the model fully trained with fair sampling (\(\alpha=1\)) has the lowest bias and drops the recall performance the most---it relatively reduces \(22.5\%\) Bias\@10 but only decreases \(10.9\%\) Recall\@10 on Flickr30K. We choose \(\alpha=0.4\) for the final model, which has a better tradeoff in retaining the recall performance. As shown in [\[fig:clip-ablation\]](#fig:clip-ablation){reference-type="ref" reference="fig:clip-ablation"}, we set the range of the clipping dimension \(m\) between \(100\) and \(400\) on MS-COCO 1K. We find that clipping too many covariates (1) harms the expressiveness of image and text representations (Recall\@1 drops from \(46.1\%\) to \(11.3\%\), Recall\@5 drops from \(75.2\%\) to \(25.4\%\), and Recall\@10 drops from \(86.0\%\) to \(34.2\%\)), and (2) causes high standard deviation in gender bias. In light of the harm on expressiveness, we select \(m=100\) for conventional use.
## Evaluation on Internet Image Search
The aforementioned evaluation results on MS-COCO and Flickr30K datasets are limited that they rely on gender labels extracted from human captions. In this sense, it is important to measure the gender biases on a benchmark where the gender labels are identified by crowd annotators. To this end, we further evaluate on the `occupation` dataset , which collects top 100 Google Image Search results for each gender-neutral occupation search term.[^7] Each image is associated with the crowd-sourced gender attribute of the participant portrayed in the image. Inspired by and, we measure the gender bias by computing the difference of expected cosine similarity between male and female occupational images. Given an occupation \(o\), the similarity bias is formulated as \[\bias = \Expectation_{v \in \mathcal{V}^o_\male} S(v, o)-\Expectation_{v \in \mathcal{V}^o_\female} S(v, o)\] where \(\mathcal{V}^o_\male\) and \(\mathcal{V}^o_\female\) are the sets of images for occupation \(o\), labeled as "male" and "female". [\[fig:occupation\]](#fig:occupation){reference-type="ref" reference="fig:occupation"} demonstrates the absolute similarity bias of CLIP and CLIP-clip on the `occupation` dataset for 18 occupations. We observe that the CLIP model exhibits severe similarity discrepancy for some occupations, including telemarketer, chemist, and housekeeper, while the *clip* algorithm alleviates this problem effectively. Note that for doctor and police officer, the CLIP-clip model exaggerates the similarity discrepancy, but the similarity bias is still less than 0.01. In general, CLIP-clip is effective for mitigating similarity bias and obtains a 42.3% lower mean absolute bias of the 100 occupations than the CLIP model (\(0.0064\) *vs.* \(0.0111\)).
# Gender Bias in Image Search
In an image search system, text queries may be either gender-neutral or gender-specific. Intuitively, when we search for a gender-neutral query like "a person is cooking", we expect a fair model returning approximately equal proportions of images depicting men and women. For gender-specific queries, an unbiased image search system is supposed to exclude images with misspecified gender information. This intention aligns with seeking more accurate search results and would be much different from the scope of measuring gender bias in gender-neutral cases. Therefore, we focus on identifying and quantifying gender bias when only searching for gender-neutral text queries.
## Problem Statement {#sec:problem-statement}
Given a text query provided by the users, the goal of an image search system is to retrieve the matching images from the curated images. In the domain of multi-modality, given the dataset \(\{(v_n,c_n)\}_{n=1}^N\) with \(N\) image-text pairs, the task of image search aims at matching every image \(v\) based on the providing text \(c\). We use \(\imageset =\{v_n\}_{n=1}^N\) to denote the image set and \(\textset = \{c_n\}_{n=1}^N\) to denote the text set. Given a text query \(c \in \textset\) and an image \(v \in \imageset\), image retrieval models often predict the similarity score \(S(v, c)\) between the image and text. One general solution is to embed the image and text into a high-dimensional representation space and compute a proper distance metric, such as Euclidean distance or cosine similarity, between vectors . We take cosine similarity for an example: \[\label{eq:similarity} \begin{split} & S(v, c) = \frac{\vec{v} \cdot \vec{c}}{\|\vec{v}\| \|\vec{c}\|} \\ \textnormal{s.t.}\ &\ \vec{v} = \mathsf{image\_encoder}(v) \\ &\ \vec{c} = \mathsf{text\_encoder}(c) \end{split}\] The image search system outputs a set of top-\(K\) retrieved images \(\mathcal{R}_{K}(c)\) with the highest similarity scores. In this work, we assume that when evaluating on test data, \(\forall c \in \mathcal{C}\), the text query \(c\) is written in gender-neutral language.
## Measuring Gender Bias {#sec:measure-gender-bias}
The situations of image search results are complex: there might be no people, one person, or more than one person in the images. Let \(g(v) \in \{\male, \female, \neutral\}\) represent the gender attribute of an image \(v\). Note that in this study gender refers to biological sex [@larson-2017-gender]. We use the following rules to determine \(g(v)\): \(g(v) = \male\) when there are only men in the image, \(g(v) = \female\) when there are only women in the image, otherwise \(g(v) = \neutral\). Portraits in image search results with different gender attributes often receive unequal exposure. Inspired by and , we measure gender bias in image search by comparing the proportions of masculine and feminine images in search results. Given the set of retrieved images \(\mathcal{R}_{K}(c)\), we count the images depicting males and females \[\begin{aligned}
N_\male & = \sum_{v \in \mathcal{R}_K(c)} \Indicator[g(v) = \male], \\ N_\female & = \sum_{v \in \mathcal{R}_K(c)} \Indicator[g(v) = \female], \end{aligned}\] and define the gender bias metric as: \[\Delta_K(c) = \begin{cases} 0, & \text{ if \(N_\male + N_\female = 0\) } \\ \frac{N_\male-N_\female}{N_\male + N_\female}, & \text{ otherwise} \end{cases}\] We don't take absolute values for measuring the direction of skewness, i.e., if \(\Delta_K(c) > 0\) it skews towards males. Note that a similar definition of gender bias \(\frac{N_\male}{N_\male + N_\female}\) in is equivalent to \((1 + \Delta(c))/2\). But our definition of gender bias considers the special case when none of the retrieved images are gender-specific, i.e., \(N_\male + N_\female = 0\). For the whole test set, we measure the mean difference over all the text queries: \[\label{eq:fairness} \bias@ K = \frac{1}{|\textset|} \sum_{c \in \textset} \Delta_K(c)\]
# Related Work {#sec:related-work}
#### Fairness in Machine Learning
A number of unfair treatments by machine learning models were reported recently, and the literature has observed a growing demand and interests in proposing defenses, including regularizing disparate impact and disparate treatment , promoting fairness through causal inference , and adding fairness guarantees in recommendations and information retrieval . The existing fair machine learning solutions can be broadly categorized as pre-processing , in-processing, and post-processing approaches. Pre-processing algorithms typically re-weight and repair the training data which captures label bias or historical discrimination .
[^1]: This study is conducted on English corpora. We will assume the text queries are all English queries hereafter.
[^2]: The data is available at [cocodataset.org](cocodataset.org).
[^3]: The data is available at <http://bryanplummer.com/Flickr30kEntities/>.
[^4]: We show the word lists in [\[app:word-list\]](#app:word-list){reference-type="ref" reference="app:word-list"}.
[^5]: The code is available at <https://github.com/kuanghuei/SCAN>.
[^6]: The pre-trained model is available at <https://github.com/openai/CLIP>.
[^7]: The data is available at <https://github.com/mjskay/gender-in-image-search>. | {'timestamp': '2021-09-14T02:19:10', 'yymm': '2109', 'arxiv_id': '2109.05433', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05433'} |
null | null |
null | null |
null | null |
null | null |
null | null |
# Introduction {#sec:intro}
Advancements in high-performance computing (HPC) and techniques like particle transport and agent-based modeling yield an enormous corpus of stochastic simulation data. provides a review of state-of-the-art simulation and modeling in such contexts as well as current challenges in this field. Simulators exist with many applications, including disease/epidemics tumor spread, inventory/supply chain management, ocean circulation, and radiation/nuclear safety. In many cases---in particular those cited above---the simulator can exhibit input-dependent-noise, or so-called heteroskedasticity. When the data is noisy, in addition to the usual nonlinear dynamics in mean structure, a large and carefully designed simulation campaign is essential for isolating signal. Modeling is another crucial component for recognizing signal in noisy data. Gaussian process (GP) regression is the canonical choice of surrogate model for simulation experiments because it provides accurate predictions and uncertainty quantification (UQ). These GP features facilitate many downstream tasks such as calibration, input sensitivity analysis, Bayesian optimization, and more. See or for a review of GPs and for additional context. However, standard GP inference and prediction scales poorly to large data sets. GP modeling involves a multivariate normal (MVN) distribution whose dimension \(N\) matches the size of the training data, i.e., \((\mathbf{X}_N, \mathbf{Y}_N)\) for regression. The quadratic cost of storage and cubic cost of decomposition of covariance matrices for determinants and inverses involved in likelihoods limits \(N\) to the small thousands. For many stochastic computer experiments in modest input dimension, small simulation campaigns are insufficient for learning whilst larger \(N\) cannot be handled by ordinary GPs. Numerous strategies abound in diverse literatures (machine learning, spatial/geostatistics, computer experiments) to scale up GP capabilities in \(N\), with many relying on approximation. Some take a divide-and-conquer approach, constructing multiple GPs in segments of a partition of the design space. A local approximate GP offers a kind of infinite-partition, separately constructing \(n \ll N\)-sized local neighborhoods around each testing location \(\mathbf{x}'\) in a transductive fashion. Small \(n\) provides thrifty processing despite \(\mathcal{O}(n^3)\) complexity. Handling multiple \(\mathbf{x}'\) is parallelizable. Predictions are accurate, but a downside to local/partition modeling is that it furnishes a pathologically discontinuous predictive surface. Other approaches target matrix decomposition directly by imposing sparsity on the covariance [@titsias2009; @aune2014parameter; @Wilson2015; @gardner2018product; @Pleiss2018; @solin2020hilbert], the precision matrix [@Datta2016; @katzfuss2021], or via reduced rank. One rank reduction strategy deploys a set of \(M \ll N\) latent knots, or *inducing points*, through which an \(\mathcal{O}(M^2N)\) decomposition is derived [e.g. @Williams2001; @Snelson2006; @Banerjee2008; @hoffman2013stochastic]. Recently, the inducing points approach has been hybridized with LAGP, improving computation times and supporting larger neighborhood sizes \(n\). GP modeling technology for computer experiments is often tailored to the no-noise (i.e., deterministic simulation) or constant (iid) noise case, which also remains true when scaling up via approximation. Partition-based schemes are an important exception, as region-based statistically independent modeling imparts a degree of non-stationarity, potentially in both mean and variance. Few mechanisms exist to limit or control partitioning. Computational independence, which speeds up the analysis by allowing multiple numerical procedures to be carried out in parallel, can be tightly coupled to regional statistical independence. LAGP is an exception, in that it allows the user to control compute time directly through the neighborhood size, \(n\). Although the software supports local inference of so-called nugget hyperparameters, its performance for noisy data is often disappointing because small \(n\) may cause noise to be misinterpreted as signal, and exacerbate predictive discontinuity. Expressly accommodating input-dependent noise has been explored in the (global/ordinary) GP literature [e.g., @goldberg1997regression; @kersting2007most], introducing \(N\) latent nugget variables which must be learned alongside the usual tunable kernel quantities. However, this strategy is in the opposite direction on the computational efficiency frontier. Stochastic kriging offers thriftier computation for GP modeling under heteroskedasticity as long as the per-run degree of replication is high. This allows latent variables to be replaced with independent moment-based variance estimates. Heteroskedastic GPs combine these two ideas: using latent variables for each of \(\bar{N}\) unique input locations and leveraging replication in design, but not requiring a minimum degree. The so-called "Woodbury identities" facilitate exact GP inference and prediction in \(\mathcal{O}(\bar{N}^3)\) time, regardless of \(N\). showed how replication, especially in high noise regions of the input space, can be essential for both statistical and computational efficiency. Yet even though \(\bar{N} \ll N\), big \(\bar{N}\) may be needed to map out the input space, which can be limiting. In this paper we propose that, by porting the Woodbury approach from HetGP over to the LIGP framework, one achieves the best of both worlds. Relatively few local-\(m\) inducing points could support a much larger number of local unique inputs \(\bar{n}\) representing a much larger neighborhood of \(n\) total observations under replication. Casting a wide \(n\)-net is crucial for separating signal from noise, but would be prohibitive under the original LAGP regime. In addition to larger neighborhoods, incorporating a state-dependent nugget in the LIGP framework, as is already available in `laGP`, is crucial in estimating noise. Local inducing points, whose number \(m \ll \bar{n} \ll n \ll N\) may be far fewer than under any of the other sizes, together with a Woodbury structure for \(\bar{n}\) representing the full \(n\), make for a powerful cascade of local information. We derive the quantities involved for inference for this upgraded LIGP, which include the gradient for numerical optimization and prediction. Illustrations are provided along the way, followed by extended benchmarking in both constant noise and heteroskedastic settings. We show that this new LIGP capability is both faster and more accurate than alternatives. The remaining flow of the paper is as follows. An overview of GP regression and LIGP is provided in Section [2](#sec:review){reference-type="ref" reference="sec:review"}. Our contribution, focused on estimating local noise and incorporating replication in LIGP via Woodbury, is detailed in Section [3](#sec:gps_with_reps){reference-type="ref" reference="sec:gps_with_reps"}. Section [4](#sec:results){reference-type="ref" reference="sec:results"} summarizes empirical work on synthetic and benchmark examples from epidemiology and biology. Section [5](#sec:bermuda){reference-type="ref" reference="sec:bermuda"} explores this new capability in the context of pricing Bermudan options in computational finance. Section [6](#sec:discussion){reference-type="ref" reference="sec:discussion"} wraps up with a discussion and directions for future work.
# Review {#sec:review}
In this section, we review standard GP surrogate modeling followed by locally induced GPs.
## Gaussian process regression {#ss:gp_basic}
Consider an unknown function \(f: \mathbb{R}^d \rightarrow \mathbb{R}\) for a set of \(d\)-dimensional design locations/inputs \(\mathbf{X}_N=(\mathbf{x}_1, \ldots,\mathbf{x}_N)^\top\) and corresponding noisy observations \(\mathbf{Y}_N=(y_1,...,y_N)^\top\) of \(f\). A common surrogate for such data is a Gaussian process (GP), which places an MVN prior on \(f\). The prior is uniquely defined by a mean vector and covariance (kernel) function \(k: \mathbb{R}^d \times \mathbb{R}^d \rightarrow \mathbb{R}\). The observation model is \(y(\mathbf{x}_i)=f(\mathbf{x}_i)+\varepsilon_i\), under constant noise, \(\varepsilon_i \stackrel{\mathrm{iid}}{\sim} \mathcal{N}(0, v)\). Often the iid noise is hyperparameterized by coupling a nugget \(g\) with scale \(\tau^2\), as in \(v = \tau^2 g\), so that the full model for responses is \[Y_N\sim \mathcal{N}_N(0, \tau^2(\mathbf{K}_N + g\mathbb{I}_N)),\] where \(\mathbb{I}_N\) denotes an \(N\times N\) identity matrix. A zero mean simplifies the exposition without loss of generality.[^1] We further assume zero-centered observations and coded/pre-scaled inputs based on inverse distance between rows of \(\mathbf{X}_N\). After pre-scaling, the entries of \(\mathbf{K}_N\) can be comprised of kernels \(k(\cdot, \cdot)\), e.g., the so-called isotropic Gaussian kernel: \[\label{eq:kernel} k_{\theta}(\mathbf{x}_i,\mathbf{x}_j)=\text{exp}\left\{ -\frac{||\mathbf{x}_i-\mathbf{x}_j||^2}{\theta}\right\}.\] A scalar lengthscale hyperparameter \(\theta\) governs the rate of radial decay of covariance. Other kernel families such as the Matérn, and so-called separable forms with lengthscales for each input coordinate are also common. Our work here is largely agnostic to such choices, as long as \(k(\cdot, \cdot)\) is differentiable in the coordinates of \(\mathbf{x}\). Many of the limiting characteristics of this overly simplistic choice (e.g., infinite smoothness and isotropy) are relaxed by a local application. Working with MVNs that involve dense covariance matrices, as is furnished by \(k(\cdot, \cdot)\), entails storage capacity that grows quadratically in \(N\). Prediction and likelihood evaluation for hyperparameter inference, such as for \((\tau^2, \theta, g)\), involve inverses and determinants requiring decomposition (usually via Cholesky) that is cubic in \(N\). For brevity, the relevant formulas are not provided here as they are not needed later. We shall provide related expressions momentarily in a more ambitious context. Special cases of the ones we provide may be found in textbooks alongside illustrations of many desirable GP properties such as excellent nonlinear predictive accuracy and UQ.
## Local approximate GPs {#ss:lagp}
Cubic computational costs are crippling in the modern context of large-\(N\) simulation experiments. As a thrifty workaround, introduced the local approximate GP (LAGP) which imposes an implicitly sparse structure by fitting a separate, limited-data GP for each predictive location \(\mathbf{x}'\) of interest. Each fit/prediction is based on an \(n \ll N\)-sized subset of data \(D_n(\mathbf{x}') = (\mathbf{X}_n(\mathbf{x}'), \mathbf{Y}_n(\mathbf{x}'))\) nearby \(\mathbf{x}'\). Multiple criteria have been suggested to build this local neighborhood \(D_n(\mathbf{x}')\), with Euclidean nearest neighbor (NN) being the simplest. Each prediction thus requires flops in \(\mathcal{O}(n^3)\), which can offer dramatic savings and is embarrassingly parallel for each \(\mathbf{x}'\). LAGP was developed for interpolating deterministic simulations, however the `laGP` software on CRAN includes a nugget-estimating capability for local smoothing. The "local nugget" feature underwhelms when applied globally for many \(\mathbf{x}' \in \mathcal{X}\) comprising a dense testing set. This can be seen first-hand by trying the examples in the `laGP` documentation. Figure [\[fig:sir_slice\]](#fig:sir_slice){reference-type="ref" reference="fig:sir_slice"} provides a demo in a stochastic simulation context with a susceptible--infected--recovered (SIR) disease model, implemented as `sirEval` in `hetGP`. We used a training set of \(\bar{N}=10000\) unique inputs in \([0,1]^2\), each paired with a random degree of replication in \(\{1,\dots,20\}\), so that the full experiment consisted of \(N \approx 100000\) runs. The views in the figure arise from a grid of 200 testing locations along the slice \(x_2=0.2\). Estimates of the "true" mean and variance (for comparison) are based on smoothed averages of 1000 replicates at each location. Comparators include: LAGP.NN with \(n=50\), LIGP.NN with \(\bar{n}=100\) (\(n\approx 1000\)) and \(m\)=10 inducing points, and HetGP on a random training subset of size \(\bar{N}=1000\). Although each comparator incoroporates a different amount of data, which may seem unfair at face value, this was necessary to balance runtime (to within an order of magnitude, in seconds: 0.5, 6.9, and 15.9 respectively) with accuracy. Observe that LAGP's estimates (blue dashed lines) of the mean and variance are highly discontinuous. The error plots in the right column show how LAGP especially struggles to produce accurate estimates for larger \(x_1\). Details on our preferred LIGP method (green dashed line), alongside further analysis of this figure, are the subject of Section [3](#sec:gps_with_reps){reference-type="ref" reference="sec:gps_with_reps"}. The problem with applying LAGP to noisy data is both fundamental and technical. Fundamentally, one needs larger \(n\) with noisy data lest the noise dominate the signal, but this severely cuts into efficiency gains. In a technical sense, many stochastic simulation experiments involve replication but the implementation does not adapt its notion of neighborhood accordingly. In the most extreme case, suppose the nearest input site to \(\mathbf{x}'\) has \(n\) replicates. Under that configuration LAGP will degenerate without a diversity of unique training inputs. Less pathologically, a multitude of nearby replicates would effectively result in a narrower \(n\)-neighborhood even though the number of sufficient statistics is fewer than \(n\). This is inefficient both statistically and computationally. Addressing these two issues is the primary contribution of this paper.
## Locally induced GPs {#ss:ligp}
LIGP is a recent extension of the LAGP idea that relaxes coupling between neighborhood construction and the local covariance structure. Rather than imposing a local covariance structure on \(Y_n(\mathbf{x}')\) directly via pairwise distances between \(\mathbf{X}_n(\mathbf{x}') \subset \mathbf{X}_N\), LIGP uses an intermediary set of \(m\) knots, or so-called *inducing points* or *pseudo inputs* \(\boldsymbol{\Psi}_m(\mathbf{x}')\), whose coordinates are not restricted to \(\mathbf{X}_N\). This yields speedups by reducing the rank of the covariance structure. When \(m \ll n\) there is a double-speedup, first through neighborhood structure (\(n \ll N\)) and then through low rank local approximation (cubic in \(m\) rather than \(n\)). Conversely, it allows a much larger neighborhood \(n\) without cubic-in-\(n\) increase in flops. Let \(\mathbf{K}_m\) denote a kernel matrix built from inducing points \(\boldsymbol{\Psi}_m(\mathbf{x}')\) and \(k(\cdot, \cdot)\), e.g., in Eq. [\[eq:kernel\]](#eq:kernel){reference-type="eqref" reference="eq:kernel"}. We defer a discussion of how one may choose \(\boldsymbol{\Psi}_m(\mathbf{x}')\) to Section [3.2](#ss:local_neighbor){reference-type="ref" reference="ss:local_neighbor"}. Similarly, write \(\mathbf{k}_{nm}\) as cross evaluations of the kernel between \(\mathbf{X}_n\) and \(\boldsymbol{\Psi}_m\). Adopting the GP approximation with inducing points introduced by yields \[\label{eq:SPGP} \mathbf{\Sigma}_n^{(m)}(\mathbf{x}') = \tau^2\left(\mathbf{k}_{nm}\mathbf{K}_m^{-1}\mathbf{k}_{nm}^\top + \Delta_n^{(m)} + g\mathbb{I}_n\right),\] as the MVN covariance deployed for inference/prediction at \(\mathbf{x}'\). Here, \(\Delta_n^{(m)}=\text{Diag}\{\mathbf{K}_n-\mathbf{k}_{nm}\mathbf{K}_m^{-1}\mathbf{k}_{nm}^\top \}\) is a diagonal correction matrix for the covariance and \(g\mathbb{I}_n\) expresses the pure noise. Our notation is suppressing some dependence on \(\mathbf{x}'\) to keep the expressions tidy. Going forward we shall use \(\Omega_n^{(m)} =\Delta_n^{(m)} + g\mathbb{I}_n\) to express the sum of diagonal matrices in Eq. [\[eq:SPGP\]](#eq:SPGP){reference-type="eqref" reference="eq:SPGP"}, electing not to embolden \(\Omega_n^{(m)}\) like we do for other matrices since it can be stored as an \(n\)-vector. When \(\boldsymbol{\Psi}_m(\mathbf{x}')\equiv \mathbf{X}_n(\mathbf{x}')\), \(\mathbf{\Sigma}_n^{(m)}\) reduces to the standard LAGP covariance \(\mathbf{\Sigma}_n(\mathbf{x}') = \tau^2 (\mathbf{K}_n+g\mathbb{I}_n)\) and offers no computational benefit. The cost savings comes through the decomposition of \(\mathbf{\Sigma}_n^{(m)}\) [\[eq:SPGP\]](#eq:SPGP){reference-type="eqref" reference="eq:SPGP"} through Woodbury matrix identities. Evaluations of \(\mathbf{\Sigma}_n^{-1(m)}\) and \(\log |\mathbf{\Sigma}_n^{(m)}|\) may be obtained with \(\mathcal{O}(m^2n)\) flops rather than \(\mathcal{O}(n^3)\). Likewise, taking \(\boldsymbol{\Psi}_m = \mathbf{X}_n = \mathbf{X}_N\) with \(\mathbf{Y}_n = \mathbf{Y}_N\) yields a global GP without approximation. Hyperparameter inference may be achieved by maximizing the logarithm of the MVN likelihood \(\mathbf{Y}_n \sim \mathcal{N}(\mathbf{0},\mathbf{\Sigma}_n^{(m)})\), which may be expressed up to an additive constant as \[\begin{aligned}
\label{eq:loglik} \ell(D_n(\mathbf{x}'),\boldsymbol{\Psi}_m;\tau^2, \theta, g) &\propto-n\log(\tau^2)-\log |\mathbf{Q}_m^{(n)}| + \log|\mathbf{K}_m|-\mathbf{1}_n^{\top}\text{log}(\Omega_n^{(m)})\mathbf{1}_n\\ & \qquad-\tau^{-2}\mathbf{Y}_n^\top\left(\Omega^{-1(m)}_n-\Omega_n^{-1(m)}\mathbf{k}_{nm}\mathbf{Q}_m^{-1(n)}\mathbf{k}_{nm}^\top\Omega_n^{-1(m)}\right)\mathbf{Y}_n. \nonumber \end{aligned}\] Above, \(\mathbf{Q}_m^{(n)}= \mathbf{K}_m + \mathbf{k}_{nm}^{\top}\Omega_n^{-1(m)} \mathbf{k}_{nm}\) and \(\mathbf{1}_n\) is a vector of \(n\) ones. Differentiating Eq. [\[eq:loglik\]](#eq:loglik){reference-type="eqref" reference="eq:loglik"} with respect to \(\tau^2\) yields a closed-form MLE \(\hat{\tau}^{2(n,m)}\), see Section [3](#sec:gps_with_reps){reference-type="ref" reference="sec:gps_with_reps"}. A numerical solver like \(\tt{optim}\) in \(\sf{R}\) can then work with negative concentrated log-likelihood (i.e., plugging \(\hat{\tau}^{2(n,m)}\) into Eq. ([\[eq:loglik\]](#eq:loglik){reference-type="ref" reference="eq:loglik"})) to estimate \(\hat{\theta}^{(n,m)}\) and \(\hat{g}^{(n,m)}\). It is important to note that did not include/estimate a local nugget \(g\), which is part of our novel contribution. For fixed hyperparameters \((\hat{\tau}^2, \hat{\theta}, \hat{g})\), a predictive distribution for \(Y(\mathbf{x}')\) arises as standard MVN conditioning via an \((n+1)\)-dimensional MVN for \((Y(\mathbf{x}'), \mathbf{Y}_n)\). Using \(\mathbf{k}_m(\mathbf{x}')=k_{\theta}(\boldsymbol{\Psi}_m,\mathbf{x}')\), the moments of that Gaussian distribution are \[\begin{aligned}
\label{eq:GPpred} \mu_{m,n}(\mathbf{x}')&=\mathbf{k}_m^{\top}(\mathbf{x}')\mathbf{Q}_m^{-1(n)}\mathbf{k}_{nm}^\top\Omega_n^{-1(m)} \mathbf{Y}_m \hspace{.5cm} \text{and}\\ \sigma_{m,n}^2(\mathbf{x}') & = \hat{\tau}^{2} \left(k_\theta(\mathbf{x}', \mathbf{x}') + \hat{g}-\mathbf{k}_m^{\top}(\mathbf{x}')\left(\mathbf{K}_m^{-1}-\mathbf{Q}_m^{-1(n)}\right)\mathbf{k}_m(\mathbf{x}')\right).\nonumber \end{aligned}\] Both the log-likelihood ([\[eq:loglik\]](#eq:loglik){reference-type="ref" reference="eq:loglik"}), and predictive equations ([\[eq:GPpred\]](#eq:GPpred){reference-type="ref" reference="eq:GPpred"}) conditional on those values, reduce to LAGP and ordinary GP counterparts with \(\boldsymbol{\Psi}_m = \mathbf{X}_n\) and \(\boldsymbol{\Psi}_m = \mathbf{X}_n = \mathbf{X}_N\) with \(\mathbf{Y}_n = \mathbf{Y}_N\), respectively. Choosing the locations of inducing points \(\boldsymbol{\Psi}_m\), globally or locally, may also be based on the likelihood, however this is fraught with issues. observed that situating \(\boldsymbol{\Psi}_m\) via classical design criteria like integrated variance, or so-called \(A\)-optimal design, led to superior out-of-sample predictive performance while also avoiding additional cubic calculation. Specifically for local analysis in LIGP, they introduced a weighted form of \(\sigma^2_{n,m}(\mathbf{x}')\), centered around \(\mathbf{x}'\), and proposed an active learning scheme that minimized weighted integrated mean-square error (wIMSE) for greedy selection of the next element \(\boldsymbol{\psi}_{m+1}\). Upgraded details are provided in Section [3](#sec:gps_with_reps){reference-type="ref" reference="sec:gps_with_reps"}. Saving on computation by stylizing the feature of the local designs for \(\boldsymbol{\Psi}_m(\mathbf{x}')\), developed several template schemes that allow to bypass optimization of wIMSE for each \(\mathbf{x}'\).
# Local smoothing under replication {#sec:gps_with_reps}
LIGP provides an opportunity to reduce compute time while also increasing neighborhood size, because complexity grows cubically in \(m\), not \(\bar{n}\) or \(n\). This is doubly important when there are replicates among \(\mathbf{X}_N\). Without inducing points, LAGP neighborhoods of size \(n\) could have very few unique inputs \(\bar{n} \ll n\). However, LIGP's Woodbury representation may be extended from inducing points to replicates so that neighborhoods of size \(\bar{n}\) may be built, regardless of how much bigger \(n\) is, without additional cubic cost.
## Fast GP inference under replication {#ss:gp_infer}
Given a local neighborhood \(D_n(\mathbf{x}')=(\mathbf{X}_n(\mathbf{x}'), \mathbf{Y}_n(\mathbf{x}'))\), let \(\bar{\mathbf{x}}_i\), \(i=1,\dots,\bar{n}\) represent the \(\bar{n} \ll n\) unique input locations. Notate each unique \(\bar{\mathbf{x}}_i\) in \(\mathbf{X}_n(\mathbf{x}')\) as having \(a_i\geq 1\) replicates, where \(\sum_{i=1}^{\bar{n}}a_i=n\). Let \(y_i^{(j)}\) be the response of the \(j^\text{th}\) replicate of \(\bar{\mathbf{x}}_i\), \(j=1,\dots,a_i\). Without loss of generality, assume the \(n\)-neighborhood inputs are ordered so that \(\mathbf{X}_n(\mathbf{x}') \equiv \mathbf{X}_n=(\bar{\mathbf{x}}_1,\dots, \bar{\mathbf{x}}_1, \dots, \bar{\mathbf{x}}_{\bar{n}})^\top\) with each unique input \(\bar{\mathbf{x}}_i\) repeated \(a_i\) times, and suppose \(\mathbf{Y}_n\) is stacked with responses from the \(a_i\) replicates in the same order. Based on this construction of \(D_n\), let \(\mathbf{X}_n=\mathbf{U}\bar{\mathbf{X}}_{\bar{n}}\), where \(\mathbf{U}\) is a \(n \times \bar{n}\) block matrix \(\mathbf{U}=\text{Diag}(\mathbf{1}_{a_1},\dots,\mathbf{1}_{a_{\bar{n}}})\). Defining \(\mathbf{U}\) in this way, we have \(\mathbf{K}_n=\mathbf{U} \mathbf{K}_{\bar{n}}\mathbf{U}^\top\) where \(\mathbf{K}_{\bar{n}} = (k_\theta(\bar{\mathbf{x}}_i,\bar{\mathbf{x}}_j))_{1\leq,i,j\leq\bar{n}}\), while \(\mathbf{U}^\top\Omega_n^{(m)}\mathbf{U}=A_{\bar{n}}\Omega^{(m)}_{\bar{n}}\) where \(A_{\bar{n}}=\text{Diag}(a_i,\dots,a_{\bar{n}})\) and \(\Omega_{\bar{n}}^{(m)} =\text{Diag}\{\mathbf{K}_{\bar{n}}-\mathbf{k}_{\bar{n}m}\mathbf{K}_m^{-1}\mathbf{k}_{\bar{n}m}^\top \} + g\mathbb{I}_{\bar{n}}\). Using these definitions, we present a new expression for \(\mathbf{\Sigma}_n^{(m)}\), extending Eq. [\[eq:SPGP\]](#eq:SPGP){reference-type="eqref" reference="eq:SPGP"} based on the \(\bar{n}\) unique input locations in the local neighborhood: \[\label{eq:SigR} \mathbf{\Sigma}_n^{(m,\bar{n})} =\tau^2(\mathbf{U}\mathbf{k}_{\bar{n}m}\mathbf{K}_m^{-1}\mathbf{k}_{\bar{n}m}^\top \mathbf{U}^\top + \text{Diag}\{\mathbf{K}_n-\mathbf{U}\mathbf{k}_{\bar{n}m}\mathbf{K}_m^{-1}\mathbf{k}_{\bar{n}m}^\top \mathbf{U}^\top \}+g\mathbb{I}_n).\] While \(\mathbf{\Sigma}_n^{(m,\bar{n})}\) in [\[eq:SigR\]](#eq:SigR){reference-type="eqref" reference="eq:SigR"} is \(n \times n\), we do not build it in practice. It is an implicit, intermediate quantity which we notate here as a step toward efficient likelihood and predictive evaluation. When \(m \ll \bar{n} \ll {n} \ll {N}\), decomposing \(\mathbf{\Sigma}_n^{(m,\bar{n})}\) to obtain \(\mathbf{\Sigma}_n^{-1(m,\bar{n})}\) and \(\log|\mathbf{\Sigma}_n^{(m,\bar{n})} |\) is much cheaper than \(\mathcal{O}(n^3)\) under the Woodbury identities. Generically, these are as follows and may be found in most matrix computation resources: \[\begin{aligned}
\label{eq:woodbury} (\mathbf{B}+\mathbf{C}\mathbf{D}\mathbf{E})^{-1}&=\mathbf{B}^{-1}-\mathbf{B}^{-1}\mathbf{C}(\mathbf{D}^{-1}+\mathbf{E}\mathbf{B}^{-1}\mathbf{C})^{-1}\mathbf{E}\mathbf{B}^{-1}\hspace{.5cm} \text{and}\\ \log|\mathbf{B}+\mathbf{C}\mathbf{D}\mathbf{E}|&=\log|\mathbf{D}^{-1}+\mathbf{E}\mathbf{B}^{-1}\mathbf{C}|+\log|\mathbf{D}|+\log|\mathbf{B}|, \nonumber \end{aligned}\] where \(\mathbf{B}\) and \(\mathbf{D}\) are invertible matrices of size \(n \times n\) and \(m \times m\) respectively, and \(\mathbf{C}^\top\) and \(\mathbf{E}\) are of size \(m \times n\). Efficient decomposition in our local approximation context involves combining LIGP's use of Eq. ([\[eq:woodbury\]](#eq:woodbury){reference-type="ref" reference="eq:woodbury"}) with inducing points, and 's () separate use leveraging replicate structure. Based on [\[eq:SigR\]](#eq:SigR){reference-type="eqref" reference="eq:SigR"}, let \[\mathbf{B}=\text{Diag}\{\mathbf{K}_n-\mathbf{U}\mathbf{k}_{\bar{n}m}\mathbf{K}_m^{-1}\mathbf{k}_{\bar{n}m}^\top \mathbf{U}^\top \}+g\mathbb{I}_n, \quad\mathbf{D}=\mathbf{K}_m^{-1} \quad \text{and }\mathbf{E}=\mathbf{C}^\top=\mathbf{k}_{\bar{n}m}^\top \mathbf{U}^\top.\] The result is as follows: \[\begin{aligned}
\label{eq:SigiR} \mathbf{\Sigma}_n^{-1(m,\bar{n})} =&\tau^{-2}\Big(\Omega_n^{-1(m)}-\mathbf{U}\Omega_{\bar{n}}^{-1(m)}\mathbf{k}_{\bar{n}m}\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^{\top}\Omega_{\bar{n}}^{-1(m)}\mathbf{U}^\top \Big) \hspace{.5cm} \text{and} \\ \log |\mathbf{\Sigma}_n^{(m,\bar{n})}|=&\log(\tau^2)+\log |\mathbf{Q}_m^{(\bar{n})}|-\log |\mathbf{K}_m| +\sum_{i=1}^{\bar{n}}a_i\text{log}\omega_i^{(\bar{n}, m)}, \nonumber \end{aligned}\] where \(\mathbf{Q}_m^{(\bar{n})}= \mathbf{K}_m + \mathbf{k}_{\bar{n}m}^{\top}\Lambda_{\bar{n}}^{(m)} \mathbf{k}_{\bar{n}m}\), \(\Lambda_{\bar{n}}^{(m)}=A_{\bar{n}}\Omega^{-1(m)}_{\bar{n}}\), and \(\omega_i^{(\bar{n}, m)}\) is the \(i^{\text{th}}\) diagonal term in \(\Omega^{(m)}_{\bar{n}}\). Eq. [\[eq:SigiR\]](#eq:SigiR){reference-type="eqref" reference="eq:SigiR"} is key to reducing computational cost from \(\mathcal{O}(nm^2)\) in Eqs. ([\[eq:loglik\]](#eq:loglik){reference-type="ref" reference="eq:loglik"}--[\[eq:GPpred\]](#eq:GPpred){reference-type="ref" reference="eq:GPpred"}) to \(\mathcal{O}(\bar{n}m^2)\). The representations in [\[eq:SigiR\]](#eq:SigiR){reference-type="eqref" reference="eq:SigiR"} allow the log-likelihood [\[eq:loglik\]](#eq:loglik){reference-type="eqref" reference="eq:loglik"} to be expressed as follows \[\begin{aligned}
\label{eq:loglikR} \ell(D_n,\boldsymbol{\Psi}_m;\tau^2, \theta, g) \propto &-n\log(\tau^2)-\log |\mathbf{Q}_m^{(\bar{n})}| + \log|\mathbf{K}_m|-\sum_{i=1}^{\bar{n}}a_i\text{log}\omega_i^{(\bar{n}, m)} \\ &-\tau^{-2}\left(\mathbf{Y}_n^\top \Omega^{-1(m)}_n\mathbf{Y}_n-\bar{\mathbf{Y}}_{\bar{n}}^\top \Lambda_{\bar{n}}^{(m)}\mathbf{k}_{\bar{n}m}\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^\top \Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}}\right), \nonumber \end{aligned}\] up to an additive constant. Above, \(\bar{\mathbf{Y}}_{\bar{n}}\) stores the averaged responses for each unique row \(\mathbf{X}_{\bar{n}}\). Notice that when Eq. [\[eq:SigiR\]](#eq:SigiR){reference-type="eqref" reference="eq:SigiR"} is applied to the log-likelihood, the \(\mathbf{U}\) terms vanish. In fact, we do not need this quantity or \(\mathbf{\Sigma}_n^{-1(m,\bar{n})}\) except as notational devices. Instead, all matrices in Eq. [\[eq:loglikR\]](#eq:loglikR){reference-type="eqref" reference="eq:loglikR"} have dimension in \(m\) and/or \(\bar{n}\). Although \(\mathbf{Y}_n^\top \Omega^{-1(m)}_n\mathbf{Y}_n\) must still be calculated, this is a product of vectors, so the entries in \(\Omega^{-1(m)}_n\) may be stored and accessed through \(\Omega_{\bar{n}}^{-1(m)}\) and \(A_{\bar{n}}\). Their storage and manipulation are thus linear in \(n\). Differentiating [\[eq:loglikR\]](#eq:loglikR){reference-type="eqref" reference="eq:loglikR"} with respect to \(\tau^2\) and solving, gives the MLE: \[\label{eq:nuhatR} \hat{\tau}^{2(\bar{n},m)} = n^{-1}\left(\mathbf{Y}_n^\top\Omega^{-1(m)}_n\mathbf{Y}_n-\bar{\mathbf{Y}}_{\bar{n}}^\top \Lambda_{\bar{n}}^{(m)}\mathbf{k}_{\bar{n}m}\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^\top \Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}}\right).\] The equation for \(\hat{\tau}^{2(\bar{n},m)}\) in [\[eq:nuhatR\]](#eq:nuhatR){reference-type="eqref" reference="eq:nuhatR"} still relies on all replicates through the full \(\mathbf{Y}_n\) and is equivalent to \(\hat{\tau}^{2(n,m)}\) as provided by (), Eq. 6. This part of the calculation is linear in \(n\). However when replicates are present, evaluation in [\[eq:nuhatR\]](#eq:nuhatR){reference-type="eqref" reference="eq:nuhatR"} may be much faster owing to the second term being sized in \(\bar{n}\) and \(m\). Plugging \(\hat{\tau}^{2(\bar{n},m)}\) into Eq. [\[eq:loglikR\]](#eq:loglikR){reference-type="eqref" reference="eq:loglikR"} yields the following concentrated negative log likelihood: \[\begin{aligned}
\label{eq:concentrate_ll} -\ell \; (D_n,\boldsymbol{\Psi}_m; \theta, g) &\propto n\log\left(\mathbf{Y}_n^\top\Omega^{-1(m)}_n\mathbf{Y}_n-\bar{\mathbf{Y}}_{\bar{n}}^\top \Lambda_{\bar{n}}^{(m)}\mathbf{k}_{\bar{n}m}\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^\top \Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}}\right)\\ &\;\; +\log |\mathbf{Q}_m^{(\bar{n})}|-\log|\mathbf{K}_m| + \sum_{i=1}^{\bar{n}}a_i\text{log}\omega_i^{(\bar{n}, m)}. \nonumber \end{aligned}\] The expression above is negated for library-based minimization in search of MLEs \(\hat{\theta}^{(\bar{n},m)}\) or \(\hat{g}^{(\bar{n},m)}\). Numerical methods such as BFGS work well. Convergence and computing time of such optimizers is aided by closed-form derivatives, which are tedious to derive but are also available in closed form: \[\begin{aligned}
\label{eq:concentrate_deriv} -\frac{\partial \ell \; (D_n,\boldsymbol{\Psi}_m, \theta, g)}{\partial \cdot} &\propto n\left(\mathbf{Y}_n^\top\Omega^{-1(m)}_n\mathbf{Y}_n-\bar{\mathbf{Y}}_{\bar{n}}^\top \Lambda_{\bar{n}}^{(m)}\mathbf{k}_{\bar{n}m}\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^\top \Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}}\right)^{-1}\\ &\quad\times \frac{\partial\left(\mathbf{Y}_n^\top\Omega^{-1(m)}_n\mathbf{Y}_n-\bar{\mathbf{Y}}_{\bar{n}}^\top \Lambda_{\bar{n}}^{(m)}\mathbf{k}_{\bar{n}m}\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^\top \Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}}\right)}{\partial \cdot}\nonumber\\ &\quad+\text{tr}\left(\mathbf{Q}_m^{-1(\bar{n})}\frac{\partial \mathbf{Q}_m^{(\bar{n})}}{\partial \cdot}\right)-\text{tr}\left(\mathbf{K}_m^{-1}\frac{\partial \mathbf{K}_m}{\partial \cdot}\right) + \sum_{i=1}^{\bar{n}} a_i \frac{\partial \log \omega_i^{(\bar{n}, m)}}{\partial \cdot}. \nonumber \end{aligned}\] Again, observe that none of these log-likelihood-derived quantities ([\[eq:loglikR\]](#eq:loglikR){reference-type="ref" reference="eq:loglikR"}--[\[eq:concentrate_deriv\]](#eq:concentrate_deriv){reference-type="ref" reference="eq:concentrate_deriv"}) involve matrices sized bigger than \(m\times \bar{n}\). Conditioning on the estimated hyperparameters leaves us with the following re-expressions of the LIGP predictive equations [\[eq:GPpred\]](#eq:GPpred){reference-type="eqref" reference="eq:GPpred"}: \[\begin{aligned}
\label{eq:GPpredR} \mu_{m,\bar{n}}(\mathbf{x}') &=\mathbf{k}_m^{\top}(\mathbf{x}')\mathbf{Q}_m^{-1(\bar{n})}\mathbf{k}_{\bar{n}m}^\top\Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}} \hspace{.5cm} \text{and}\\ \sigma_{m,\bar{n}}^2(\mathbf{x}') & =\hat{\tau}^{2(\bar{n}, m)} \left(k_\theta(\mathbf{x}',\mathbf{x}')+\hat{g}^{(\bar{n}, m)}-\mathbf{k}_m^{\top}(\mathbf{x}')\left(\!\mathbf{K}_m^{-1}-\mathbf{Q}_m^{-1(\bar{n})} \right)\mathbf{k}_m(\mathbf{x}') \right).\nonumber \end{aligned}\] Although these look superficially similar to [\[eq:GPpred\]](#eq:GPpred){reference-type="eqref" reference="eq:GPpred"}, the following remarks are noteworthy. Perhaps the biggest difference is that the average of replicates \(\bar{\mathbf{Y}}_{\bar{n}}\) is used for the mean. The Woodbury identity ensures that the result is the same as [\[eq:GPpred\]](#eq:GPpred){reference-type="eqref" reference="eq:GPpred"}, both in mean and variance, if the replicate structure is overlooked. In both cases, many of the details are buried in matrices whose \((\bar{n})\) scripts mask a substantial difference in the subroutines that would be required to build the requisite quantities for [\[eq:GPpredR\]](#eq:GPpredR){reference-type="eqref" reference="eq:GPpredR"} as compared to [\[eq:GPpred\]](#eq:GPpred){reference-type="eqref" reference="eq:GPpred"}. Kriging with pre-averaged responses \(\bar{\mathbf{Y}}_{\bar{n}}\), whether locally or globally, is a common technique for efficiently managing replicates. However, one must take care to (a) adjust the covariance structure appropriately, and (b) utilize all replicates in estimating scale \(\hat{\tau}^2\). show that \(\bar{\mathbf{Y}}_{\bar{n}}\) are not sufficient statistics. Using only pre-averaged responses risks leaving substantial uncertainty un-accounted for in prediction. Our Woodbury identities provide a full covariance structure ([\[eq:SigR\]](#eq:SigR){reference-type="ref" reference="eq:SigR"}), and the adjustments in \(n\)-quantities for an appropriate scale estimate ([\[eq:nuhatR\]](#eq:nuhatR){reference-type="ref" reference="eq:nuhatR"}). Alternative approaches, such as 's stochastic kriging (SK), also use unique \(\bar{n}\) inputs and pre-averaged outputs. Asymptotically covering (a), it may be shown that SK yields the best linear unbiased predictor (BLUP) when conditioning on estimated hyperparameters. However, the uncertainty in that predictive surface is not fully quantified. For example, SK may only furnish confidence intervals on predictions, not full predictive intervals, because it does not utilize the full set of sufficient statistics (b). Since our Woodbury application links Eq. [\[eq:GPpredR\]](#eq:GPpredR){reference-type="eqref" reference="eq:GPpredR"} with [\[eq:GPpred\]](#eq:GPpred){reference-type="eqref" reference="eq:GPpred"} identically, a BLUP property holds (locally) for LIGP without asymptotic arguments, and with full predictive UQ. More specifically, consider the following comparison which upgrades an argument from to our local inducing points setting. We show how the MLE of the scale parameter \(\tau^2\) is not the same (conditioned on \(\theta, g\)) when calculated on the original set of data versus the unique design locations with averaged responses. Using unique-\(\bar{n}\) calculations based on \(\bar{\mathbf{Y}}_{\bar{n}}\) only would result in \(\bar{\tau}^{2(\bar{n},m)}=\bar{n}^{-1}\bar{\mathbf{Y}}_{\bar{n}}^\top \bar{\boldsymbol{\Sigma}}_{\bar{n}}^{-1}\bar{\mathbf{Y}}_{\bar{n}}\) where \[\label{eq:Sig_avg} \bar{\boldsymbol{\Sigma}}_{\bar{n}}^{(m)}=\mathbf{k}_{\bar{n}m}\mathbf{K}_m^{-1}\mathbf{k}_{\bar{n}m}^\top+\Delta_{\bar{n}}^{(m)} +gA_{\bar{n}}^{-1},\] weights the nugget (noise) based on the number of replicates \(a_i\) at each unique location \(\bar{\mathbf{x}}_i\). Our full neighborhood calculation for the local MLE for \(\tau^{2}\) with inducing points in Eq. [\[eq:nuhatR\]](#eq:nuhatR){reference-type="eqref" reference="eq:nuhatR"} can be rewritten as follows: \[\begin{aligned}
\label{eq:tau2_avg} \hat{\tau}^{2(\bar{n},m)} &=N^{-1}\biggl(\bar{n}\bar{\tau}^{2(\bar{n},m)} + \mathbf{Y}_n^\top\Omega^{-1(m)}_n\mathbf{Y}_n-\bar{\mathbf{Y}}_{\bar{n}}^\top \Lambda_{\bar{n}}^{(m)}\bar{\mathbf{Y}}_{\bar{n}}-\\ & \qquad\qquad \bar{\mathbf{Y}}_{\bar{n}}^\top \bar{\boldsymbol{\Sigma}}_{\bar{n}}^{-1(m)}((A_{\bar{n}}^{-1}-\mathbb{I}_{\bar{n}})^{-1}\Delta_{\bar{n}}^{-1(m)}+\bar{\boldsymbol{\Sigma}}_{\bar{n}}^{-1(m)})^{-1}\bar{\boldsymbol{\Sigma}}_{\bar{n}}^{-1(m)}\bar{\mathbf{Y}}_{\bar{n}}\biggr). \nonumber \end{aligned}\] See the Appendix for details. Observe that \(\bar{\tau}^{2(\bar{n},m)}\) above in Eq. [\[eq:Sig_avg\]](#eq:Sig_avg){reference-type="eqref" reference="eq:Sig_avg"} is but a small part of [\[eq:tau2_avg\]](#eq:tau2_avg){reference-type="eqref" reference="eq:tau2_avg"}. The term following \(\bar{\tau}^{2(\bar{n},m)}\) serves as a correction for the variance estimate.
## Local neighborhood geography {#ss:local_neighbor}
To dig a little deeper into the differences between LAGP and LIGP, especially regarding handling replicates and LIGP's upgraded capability, consider again the SIR example introduced in Section [2.2](#ss:lagp){reference-type="ref" reference="ss:lagp"}. Whereas Figure [\[fig:sir_slice\]](#fig:sir_slice){reference-type="ref" reference="fig:sir_slice"} involved a continuum of predictions along a 1d slice, Figure [\[fig:sir_neighborhood\]](#fig:sir_neighborhood){reference-type="ref" reference="fig:sir_neighborhood"} shows relevant quantities involved in one of the locations \(\mathbf{x}'=(0.47, 0.2)\) along that slice, indicated by the green triangle. Recall that the input space is in 2d with \(\bar{N}=10000\) unique inputs and replicate degree \(a_i\in \{1, 2, \dots, 20\}\) distributed uniformly. With such a high degree of replication, LAGP's default \(n=50\) neighborhood is small, comprising of only \(\bar{n}=5\) unique locations. As the green triangle moves along the slice of Figure [\[fig:sir_slice\]](#fig:sir_slice){reference-type="ref" reference="fig:sir_slice"}, eventually one of those five will be replaced, resulting in a change in conditioning set of upwards of 1/5 of \(\bar{n}\). This provides some of the intuition as to why the LAGP surfaces in Figure [\[fig:sir_slice\]](#fig:sir_slice){reference-type="ref" reference="fig:sir_slice"} are so "jumpy" and often inaccurate. LIGP's implementation of Eqs. ([\[eq:loglikR\]](#eq:loglikR){reference-type="ref" reference="eq:loglikR"}--[\[eq:GPpredR\]](#eq:GPpredR){reference-type="ref" reference="eq:GPpredR"}) affords a much larger amount of data in the local model for commensurate computing effort through a template of \(m=10\) inducing points. These are indicated as blue triangles in Figure [\[fig:sir_neighborhood\]](#fig:sir_neighborhood){reference-type="ref" reference="fig:sir_neighborhood"}. The neighborhood of \(\bar{n}=100\) unique design locations are indicated with their number of replicates (black numbers), \(n=1091\) in total. In this instance, LIGP conditions upon 20 times more training data than LAGP. This yields more accurate and stable mean and variance estimates, shown visually along a slice in Figure [\[fig:sir_slice\]](#fig:sir_slice){reference-type="ref" reference="fig:sir_slice"} and summarized in a more expansive Monte Carlo (MC) exercise in Section [4.3](#ss:sir){reference-type="ref" reference="ss:sir"}. Those latter results additionally show that such accurate predictions can be achieved at a computational cost that is substantially lower than LAGP. The location and spacing of the inducing points \(\boldsymbol{\Psi}_m(\mathbf{x}')\) in Figure [\[fig:sir_neighborhood\]](#fig:sir_neighborhood){reference-type="ref" reference="fig:sir_neighborhood"} (blue triangles) come from a local space-filling template scheme based on wIMSE. The idea is that a local inducing point set of a desired size can be built up greedily, using integrated variance under a Gaussian measure centered at \(\mathbf{x}'\). An expression using \(\bar{n}\) unique neighborhood elements (together with all replicates, \(n\)) given [\[eq:SigR\]](#eq:SigR){reference-type="eqref" reference="eq:SigR"} is \[\begin{aligned}
\label{eq:wimseR} &\text{wIMSE}_{\bar{n}}^{(m+1)} (\boldsymbol{\psi}_{m+1}, \mathbf{x}') =\int_{\breve{\mathbf{x}}\in \mathcal{X}}k_\theta(\breve{\mathbf{x}},\mathbf{x}')\frac{\sigma_{m+1,\bar{n}}^2(\breve{\mathbf{x}})}{\tau^2} \; d\breve{\mathbf{x}} \\ & \quad =\frac{\sqrt{\theta\pi}}{2}\prod_{k=1}^d\Bigg(\text{erf}\left\{\frac{\mathbf{x}'-a_k}{\sqrt{\theta}}\right\}-\text{erf}\left\{\frac{\mathbf{x}'-b_k}{\sqrt{\theta}}\right\}\Bigg)-\text{tr}\Big\{\Big(\mathbf{K}^{-1}_{m+1}-\mathbf{Q}^{-1(\bar{n})}_{m+1}\Big)\mathbf{W}'_{m+1}\Big\}, \nonumber \end{aligned}\] where \(\text{erf}\) is the Gaussian error function, \(a_k,b_k\) are bounds for a hyperrectangle study region \(\mathcal{X} = [a_k, b_k]_{k=1}^d\), and \(\mathbf{W}'_{m+1}=\prod_{k=1}^d \mathbf{W}'_{m+1,k}\). Again, the resemblance to a similar expression from masks a substantial enhancement in its component parts due to a double-Woodbury application, as notated by \(\bar{n}\) scripts. The \((i,j)^\text{th}\) entry of \(\mathbf{W}'_{m+1,k}\) is \(w_{m+1,k}(\boldsymbol{\psi}_{i}, \boldsymbol{\psi}_{j})\) defined in Eq. (11) in. A closed-form gradient with respect to \(\boldsymbol{\psi}_{m+1}\) may also be derived. This is similar in form to, so we do not duplicate it here. Given \(\mathbf{Q}^{-1(\bar{n})}_{m+1} \equiv \mathbf{Q}^{-1(n)}_{m+1}\), the calculation in [\[eq:wimseR\]](#eq:wimseR){reference-type="eqref" reference="eq:wimseR"} and derivative are equivalent to Eqs. (10--12) in that paper. Consequently, a wIMSE optimized design of inducing points is not affected by the local distribution of replicates. Under replication \(\mathbf{Q}^{-1(\bar{n})}_{m+1}\) can, of course, be calculated much faster than \(\mathbf{Q}^{-1(n)}_{m+1}\), so the optimization is speedier. Regularity in local inducing point design also means that simple space-filling designs can serve as thrifty substitutes to the wIMSE design with similar predictive accuracy. The qNorm method transforms a space-filling design with the inverse-CDF of a Normal distribution to mimic an wIMSE inducing point design. This is what was used for Figures [\[fig:sir_slice\]](#fig:sir_slice){reference-type="ref" reference="fig:sir_slice"}--[\[fig:sir_neighborhood\]](#fig:sir_neighborhood){reference-type="ref" reference="fig:sir_neighborhood"}. Once a set of inducing points \(\boldsymbol \Psi_m(\mathbf{x}')\) is calculated for one \(\mathbf{x}'\), show that it can serve as a *template* for predictions at other members of a testing set \(\mathcal{X}\) through displacement and warping, so long as the characteristics of the original design \(\mathbf{X}_N\) are uniform. Templates based on wIMSE tend to fill the neighborhood in a more uniform manner than via qNorm, providing a higher density of inducing points near the predictive location. This often leads to more accurate prediction. Yet the reduced time required to build qNorm templates is also beneficial. We contrast both options in our empirical work in Section [4](#sec:results){reference-type="ref" reference="sec:results"}.
# Implementation and benchmarking {#sec:results}
Now we provide practical details and report on experimental results showcasing LIGP's potential for superior accuracy and UQ at lower computational costs against LAGP and HetGP. Our examples include data from a toy function and real stochastic simulators. All analysis was performed on an eight-core hyperthreaded Intel i9-9900K CPU at 3.60 GHz. Every effort was made to take advantage of that distributed resource to minimize compute times.
## Implementation details {#ss:implement}
Open source implementation of LIGP can be found in the `liGP` package on CRAN. R code supporting all examples reported here and throughout the paper, may be found on our Git repository.
Our main competitors are `laGP` and `hetGP`. While `laGP` is coded in C with `OpenMP` for symmetric multiprocessing (SMC) parallelization (R serving only as wrapper), and `hetGP` leverages substantial `RCpp`, our LIGP implementation coded purely in R using `foreach` for SMC distribution. With `laGP` we use the default neighborhood size of \(n=50\) built by NN and Active Learning Cohn, representing distance-and variance-based (similar to wIMSE) approaches, respectively. For `hetGP`, we reduce the training data to a random subset of \(\bar{N}=1000\) unique inputs (retaining all replicates) to keep decompositions tractable. Despite compiled C/C++ libraries and thrifty (local/global) data subset choices, our (more accurate and larger-neighborhood) LIGP models are competitive, time-wise, and sometimes notably faster. Our LIGP implementation uses an isotropic Gaussian kernel with scalar lengthscale \(\theta\). To improve numerical conditioning of matrices \(\mathbf{K}_m\) and \(\mathbf{Q}_m^{(n)}\) for stable inversion, we augment their diagonals with \(\epsilon_K=10^{-8}\) and \(\epsilon_Q=10^{-5}\) jitter, respectively. Our implementation in `liGP` automatically increases these values for a particular local model if needed. LAGP and HetGP results also utilize an isotropic Gaussian kernel. Using separable local formulations do not significantly improve predictive performance, especially after globally prescaling the inputs. Pre-scaling or warping of inputs has become a popular means of boosting predictive performance of sparse GPs. In our exercises, we divide by square-rooted separable global lengthscales obtained from a GP fit to random \(\bar{N}=1000\) data subsets with averaged responses \(\bar{\mathbf{Y}}_{\bar{N}}\). See, Section 9.3.4, for details. The time required for this, which is negligible relative to other timings, is not included in our summaries. For inducing point designs \(\boldsymbol{\Psi}_m(\mathbf{x}')\), we use wIMSE templates based on Eq. [\[eq:wimseR\]](#eq:wimseR){reference-type="eqref" reference="eq:wimseR"} built at the center of the design space \(\check{\mathbf{x}}\), determined as the median of \(\mathbf{X}_N\) in each dimension. For initial local lengthscale \(\theta^{(0)}\), we adopt a strategy from `laGP` via the 10% quantile of squared pairwise distances between members of \(\mathbf{X}_n\).[^2] When selecting each \(\boldsymbol{\psi}_{m+1}\) to augment \(\boldsymbol{\Psi}_m\), optimizing wIMSE is conducted via a 20-point multi-start derivative-based L-BFGS-B scheme (using `optim` in R to a tolerance of 0.01) peppered within the bounding box surrounding the neighborhood \(\mathbf{X}_n\). In addition to the wIMSE design for the inducing point template, we consider so-called "qNorm" templates derived from space-filling designs. These originate from \(m-1\) point Latin hypercube samples on the hyperrectangle enclosing \(\mathbf{X}_n(\check{\mathbf{x}})\), augmented with \(\check{\mathbf{x}}\) as the \(m^\mathrm{th}\) point. These LHSs are warped with the inverse Gaussian CDF to closely mimic the wIMSE design. Individual local neighborhoods and predictions are made for a set of \(N'\) testing locations \(\mathbf{x}' \in \mathcal{X}\) given training data \(\{\mathbf{X}_N,\mathbf{Y}_N\}\), neighborhood size \(\bar{n}\), and number of inducing points \(m\). We use \(\bar{n}=100\) for each experiment, with \(m=10\) and \(30\) for the 2d and 4d problems, respectively. Each location \(\mathbf{x}'_i\), for \(i = 1, \dots, N'\) proceeds in parallel via 16 `foreach` threads.[^3] To estimate scale and lengthscale, we used Eqs. ([\[eq:nuhatR\]](#eq:nuhatR){reference-type="ref" reference="eq:nuhatR"}--[\[eq:concentrate_deriv\]](#eq:concentrate_deriv){reference-type="ref" reference="eq:concentrate_deriv"}) via `optim` through the local neighborhoods of \(\mathbf{x}'\). To aid in discerning signal from noise during hyperparameter optimization, we incorporate default priors for \(\theta\) and \(g\), from `laGP`, described in Appendix A of. Predictions follow Eq. [\[eq:GPpredR\]](#eq:GPpredR){reference-type="eqref" reference="eq:GPpredR"}.
## Benchmark non-stationary data {#ss:syn_results}
The toy 2d function known as Herbie's tooth is attractive as a benchmark problem due to its non-stationary mean surface with multiple local minima. The mean function is defined by \(f(x_1, x_2)=-w(x_1)w(x_2)\), \(x_1,x_2 \in [-2,2]\), where \[w(x)=\text{exp}\left\{-(x-1)^2\right\}+\text{exp}\left\{-0.8(x+1)^2\right\}-0.05\sin\left(8(x+0.1)\right).\] For this experiment we introduce constant noise \(\epsilon \sim N(0,0.02^2)\), creating a response \(y(x_1, x_2) = f(x_1, x_2) + \epsilon\). Each of the 30 MC repetitions is comprised of a fresh training set of \(\bar{N}=10000\) LHS locations, each with \(a_i \stackrel{\mathrm{iid}}{\sim} \mathrm{Unif}\{1,2,\dots, 20\}\) for \(i=1,\dots,\bar{N}\). Although our later experiments use a more common, fixed-degree setup for replicates, we choose random \(a_i\) here to underscore that LIGP does not require a minimum amount of replication. A separate LHS of \(N'=10000\) out-of-sample testing locations is created for each MC run. Our LAGP fits include \(n=100\) via NN for a slightly fairer comparison to LIGP, in addition to the other LAGP options described earlier. Figure [\[fig:ht_results\]](#fig:ht_results){reference-type="ref" reference="fig:ht_results"} shows out-of-sample accuracy statistics versus the logarithm of compute time across the MC repetitions. A particular method's median statistic, taken marginally across both axes, is represented with a bold symbol. The left panel shows the root mean-squared prediction error (RMSE; lower is better) against the (de-noised) truth along the vertical axis. In the right panel, the vertical axis shows proper scores (higher is better), combining mean and variance (UQ) accuracy against the (noisy) test data. Concretely, RMSE (lower is better) and score (higher) follow: \[\label{eq:rmse_score} \text{RMSE}=\sqrt{\sum_{i=1}^{N'}(\hat{\mu}(\mathbf{x}_i)-y(\mathbf{x}_i))^2} \hspace{.4cm} \text{and} \hspace{.4cm} \text{Score}=-\sum_{i=1}^{N'}\frac{\left(\hat{\mu}(\mathbf{x}_i)-y(\mathbf{x}_i)\right)^2}{\hat{\sigma}^2(\mathbf{x}_i)}-\sum_{i=1}^{N'}\hat{\sigma}^2(\mathbf{x}_i).\] LIGP methods yield the fastest predictions with the lowest RMSEs and highest scores. Our LIGP.qNorm (blue \(+\)'s) has the quickest compute time. LAGP.NN with \(n=50\) (green inverted triangles) is the fastest among the LAGP models, but is least accurate. HetGP (black circles) is fit with substantial sub-setting, which explains its high-variance metric on both time and accuracy. Notice that LAGP.ALC (red diamonds) and LAGP.NN (\(n=100\); navy triangles) perform similarly for RMSE, but the latter has a higher score. This supports our claim that a wider net of local training data is needed to better estimate noise.
## Susceptible-infected-recovered (SIR) epidemic model {#ss:sir}
We return to the SIR model first mentioned in Section [2.2](#ss:lagp){reference-type="ref" reference="ss:lagp"}. SIR models are commonly used for cost-benefit analysis of public health measures to combat the spread of communicable diseases such as influenza or Ebola. We look at a simple model with two inputs: \(x_1\) the initial number susceptible individuals; and \(x_2\) the number of infected individuals. In `hetGP`, the function `sirEval` accepts these two inputs on the unit scale. The response of the model is the expected aggregate number of infected-days across Markov chain trajectories until the epidemic ends. The signal-to-noise ratio varies drastically throughout \(\mathcal{X}\), making this model an ideal test problem for LIGP. In this experiment we keep the same \(\bar{N}=10000\) training set size as Section [4.2](#ss:syn_results){reference-type="ref" reference="ss:syn_results"}, but now use a fixed degree of 10 replicates at each location, a more common default design setup in the absence of external information about regions of high/low variability. We also keep a large \(N'=10000\) testing set, which helps manage MC error in our out-of-sample metrics in this heteroskedastic setting. The experiment's results are displayed in Figure [\[fig:sir_results\]](#fig:sir_results){reference-type="ref" reference="fig:sir_results"}. The left panel's vertical axis shows the RMSE values between each method's mean predictions and the noisy testing observations. Variability across repetitions in the data generating mechanism looms large compared to the differences in RMSE values among the methods. Although RMSE/scores may look similar marginally, within each MC repetition there is a clearer ordering. To expose that, the left section of Figure [\[fig:sir_results\]](#fig:sir_results){reference-type="ref" reference="fig:sir_results"}'s table ranks the methods based on their RMSE (lowest/best first). The reported \(p\)-values are for one-sided paired Wilcoxon tests between a method's RMSE and the RMSE from the next best fit. Observe, for example, that LIGP.qNorm (blue \(+\)'s) produces a significantly lower distribution of RMSEs compared to the other models, with LIGP.wIMSE (purple \(\times\)'s) and HetGP (black circles) not far behind. Score (right panel) offers clearer distinctions between the methods, with LIGP besting LAGP and HetGP via paired Wilcoxon test. Using score as a metric highlights the differences in accuracy for the model estimates. By weighting the squared error by the predictive variance, not all errors of the same magnitude are considered equal (unlike in RMSE), which is crucial when modeling a heteroskedastic process. LIGP.qNorm and LIGP.wIMSE are statistically distinguishable from each other (\(p\)-value \(< 0.01\)) and better than LAGP.NN with \(n=100\) (\(p\)-value \(< 10^{-9}\)), the next in the list. There is more variation in the computation times of the LIGP models compared to LAGP, which is likely due to LIGP's tendency to need more function evaluations for hyperparameter optimization. LIGP provides significant time savings despite an R implementation.
## Ocean oxygen concentration model {#ss:fksim}
Here we consider a stochastic simulator modeling oxygen concentration deep in the ocean. This highly heteroskedastic simulator uses MC to approximate the solution of an advection-diffusion equation from computational fluid dynamics. There are four inputs: latitude and longitude coordinates, and two diffusion coefficients. The code required to run the simulation can be found in our repository. Inputs are scaled to the unit cube \([0,1]^4\); the output is the ocean oxygen concentration. We keep the same training/testing set sizes from the SIR experiment: \(\bar{N}=10000\) with 10 replicates each and \(N'=10000\). Figure [\[fig:fksim_results\]](#fig:fksim_results){reference-type="ref" reference="fig:fksim_results"} shows out-of-sample metrics similarly mirroring that experiment. At first glance, HetGP (black circles) is clearly worse (higher RMSEs, lower scores) than LIGP/LAGP. In 4d, using a subset of only 1000 unique locations does not provide enough information to model this surface. The variation in compute time for HetGP, due to challenges in modeling the latent noise, support this claim. Focusing on LIGP and LAGP results, we find LIGP again among the best in RMSE, score, and computation time. LAGP.NN with the larger neighborhood of \(n=100\) (navy triangle) does well at modeling the mean and noise surfaces (its RMSE and score medians are comparable to the LIGP models), but takes nearly 6 times longer than LIGP.qNorm. HetGP (black circles) and LAGP.ALC (red diamonds) model the mean surface well, producing slightly worse RMSEs, but struggle with UQ as measured by proper score.
# Bermudan Option Pricing {#sec:bermuda}
An application of high-throughput stochastic simulation arises in computational methods for optimal stopping, which are particularly motivated by quantitative finance contexts. Several types of financial contracts--the so-called American-type claims--offer their buyer the option to exercise, i.e., collect, her contract payoff at any time until its maturity. Pricing and risk managing such a contract requires analysis of the respective optimal exercise strategy, such deciding when to exercise the option. Monte-Carlo-driven valuation of American-type claims consists of applying the dynamic programming formulation to divide the global dynamic control problem into local single-step optimization sub-problems. The latter revolve around partitioning the input space (interpreted as the underlying prices at the current time-step) into the *exercise* and *stopping regions*, to be learned by constructing a certain statistical surrogate. More precisely, one wishes to compare the \(q\)-value, namely the expected future payoff conditional on not exercising now, with the immediate payoff. While the latter is a known function, the \(q\)-value is specified abstractly as a recursive conditional expectation and must be learned. In the literature, it has been demonstrated that GP surrogates are well suited for this emulation task, due to their flexibility, few tuning parameters, and synergy with sequential design heuristics. Nevertheless, their cubic scaling in design size \(N\) is a serious limitation. The nature of the financial context implies an input space is 2-5 dimensional (1d problems are also common but are computationally straightforward) and a very low signal-to-noise ratio. Consequently, Monte Carlo methods for accurate estimation of the \(q\)-value call for \(N \gg 10^4\) simulations. demonstrated that replication is pivotal to addressing this scaling challenge; explored adaptive replication. LIGP offers an additional boost by leveraging replication and local approximation, simultaneously addressing speed, accuracy and localization. Relative to existing surrogates LIGP brings several improvements. First, it overcomes limitations on the number of training simulations \(N\). In previous experiments working with matrices of size \(\gg 2000\) was prohibitively slow. This limited traditional GPs to data with dimension \(d \le 3\) or excessive replication. Second, LIGP organically handles replicated designs that are needed to separate signal from noise, which is heteroskedastic and non-Gaussian in this application. Third, the localization intrinsic in LIGP is beneficial to capture non-stationarity. Typically the \(q\)-value is flat at the edges and has a strong curvature in the middle of the input space. Finally, the control paradigm implies that the surrogate is intrinsically prediction-driven--the main use-case being to generate exercise rules along a set of price scenarios. Thus, the surrogate is to be evaluated at a large number of predictive locations and LIGP offers trivial parallelization that vastly reduces compute time relative to plain GP variants.
## Illustrating the Exercise Strategy
We have added LIGP regressors as a new choice to the `mlOSP` library for R, which contains a suite of test cases that focus on valuation of American-type options on multiple assets. This allows benchmarking the LIGP module against alternatives, including plain GP and `hetGP` solvers that are already part of `mlOSP`. To fix ideas, we consider the \(d\)-dimensional Bermudan Max-Call with payoff function \[h_\mathrm{MaxCall}(k,\mathbf{x}) = e^{-r k\Delta t}( \max_{i \le d} x_i-{\cal K})_+.\] The parameter \({\cal K}\) is known as the strike, \(r\) is the continuously compounded interest rate and \(\Delta t\) is the exercise frequency measured in calendar time (e.g. daily, \(\Delta t = 1/365\)). The input space \(\mathbf{x} \in \mathbb{R}^d_+\) represents prices \(x_1, \ldots, x_d\) of \(d\) stocks. Thus, the buyer of the Max-Call is able to collect \(h_\mathrm{MaxCall}(k,\mathbf{X}(k\Delta t))\) dollars, the difference between the largest stock price and the strike, at any one of the \(K\) exercise opportunities, \(k=1,\ldots, K\). The choice of when to do so leads to \(K\) decisions regarding whether to exercise the option at time step \(k\) (assuming it was not exercised yet) or not. Thus, we need \(K-1\) surrogates for the \(q\)-value \(q(k,\mathbf{x})\), \(k=0,\ldots,K-1\). The underlying stochastic simulator is based on generating trajectories of the asset prices. We stick to independent log-normal (Geometric Brownian motion) dynamics for the price of the \(i^{\mathrm{th}}\) asset, \[\begin{aligned}
\label{eq:gbm} dX_i(t) = (r-\delta_i) X_i(t) dt + \sigma_i X_i(t) dW_i(t), \qquad i=1,\ldots, d, \end{aligned}\] where \(W_i(t)\) are independent standard Brownian motions, \(\sigma_i\) is the volatility, and \(\delta_i\) the continuous dividend yield of the \(i\)-th asset. This implies that given \(X_i(s)\) for any \(t>s\), \(X_i(t)\) has a log-normal distribution. Denote by \(\mathbf{X}_{k:K} = \mathbf{X}_{k\Delta t}, \ldots, \mathbf{X}_{K \Delta t}\) a realized trajectory of the \(d\) asset prices on the interval \([k\Delta t, K\Delta t]\). The stochastic simulator yields empirical samples of \(H_k(\mathbf{X}_k, \mathbf{X}_{k+1}, \ldots, \mathbf{X}_K)\) where the aggregate function \(H_k\), representing the pathwise payoff, is a selector \[H_k(\mathbf{x}_k,\ldots,\mathbf{x}_K) = \left\{\sum_{s=k+1}^{K} h(s, \mathbf{x}_{s}) 1_{A_s}\right\}-h(k, \mathbf{x}_k),\] and \(A_k\) records the first pathwise exercise time along the given trajectory. The objective of learning the \(q\)-value at step \(k\) is equivalent to regressing \(H_k\) against \(\mathbf{x}_k\) (exploiting the Markov property of the state dynamics) to obtain the so-called *timing value* \(\hat{T}_k(\mathbf{x}_k)\) and identify regions where the conditional mean is positive/negative. Note that this task must be done for *each* time step \(k=K-1,\ldots, 1\), and by construction the outputs at step \(k\) depend on the \(\hat{T}_{k+1:K-1}\)'s fitted previously. This iterative estimation leads to non-trivial error back-propagation along the financial time dimension in \(k\). Returning to the LIGP implementation, we have a sequence of surrogates that must be constructed. For each surrogate, we are given a data set of size \(N\) with \(\bar{N}\) unique inputs in \(\mathbb{R}^d_+\) and a constant number of replicates \(a\), \(N = \bar{N} \cdot a\). For the experimental design we take an LHS on a user-specified sub-domain. After fitting the surrogate for \(\hat{T}_k\), we need to predict on \(N'=N\) unique outputs, which lie in the training sub-domain but are otherwise distinct from the training inputs. To illustrate how that looks, Figure [\[fig:bermudan2d\]](#fig:bermudan2d){reference-type="ref" reference="fig:bermudan2d"} compares global GP and LIGP fits for the timing value of the 2d Bermudan Max-Call option. The model parameters are \({\cal K}=100\), \(r=0.05\), \(\delta_1=\delta_2=0.1\), \(\sigma_{1}=\sigma_{2} =0.2\), \(T=3\) and \(\Delta t = 1/3\) so that there are \(K=9\) time periods. We display the fitted surrogates at one intermediate time step, \(k=6\). In this problem, the option is out-of-the-money when both \(x_1\) and \(x_2\) are below the strike \({\cal K}=100\), hence the exercise strategy is to continue in that region and the latter (the white square in the bottom-left of each panel) is excluded from the surrogate prediction. For both problems, we have a randomized training set of \(\bar{N} \approx 650\) inputs, with \(a=25\) replicates each, created by a LHS design. The LIGP surrogate uses \(m=10,n=50\), with estimated nugget and fixed lengthscale \(\theta =1\). The estimated strategy is to continue when the timing value \(\hat{T}_k\) (color coded in both panels) is positive and to stop when it is negative, with the zero-contour (in red) demarcating the boundary of the stopping region. The stopping region consists of two disconnected components, in the bottom-right and upper-left; due to the symmetric choice of parameters, those regions (and the entire response) are symmetric along the \(x_1=x_2\) diagonal. Thus, the response surface features a diagonal ridge with a global maximum around \((100,100)\) and two basins on either side. The height of the ridge and the depth of the basins change in \(k\). The respective simulation variance is low in regions where \(T_k(\mathbf{x}) \simeq 0\) and is high in regions where \(|T_k(\mathbf{x})|\) is large. We emphasize that Figure [\[fig:bermudan2d\]](#fig:bermudan2d){reference-type="ref" reference="fig:bermudan2d"} is purely diagnostic; actual performance of the surrogate is based on the resulting estimated option price. The latter is obtained by generating a test set of \(N'\) scenarios and using the collection of \(K\) surrogates \(\hat{T}_k\), \(k=0,\ldots, K-1\) to evaluate the resulting exercise times and ultimately the *average* payoff. Fixing a test set, the surrogate that yields a higher expected payoff is deemed better. Traditionally, a test set is obtained by fixing an initial \(\mathbf{X}(0)\) and sampling \(N'\) i.i.d. paths of \(\mathbf{X}_{0:K}\). For example, taking \(\mathbf{X}(0) = (90,90)\) and using the surrogates displayed in Figure [\[fig:bermudan2d\]](#fig:bermudan2d){reference-type="ref" reference="fig:bermudan2d"} we obtain an estimated Max-Call option value of 8.02 via LIGP and 8.00 via the global GP. Note that this financial assessment of surrogate accuracy is not based on standard IMSE-like metrics, but is driven by the accuracy of correctly learning the exercise strategy, namely the zero-contour of \(H_k\). Hence, the goal is to correctly specify the *sign* of \(T_k(\mathbf{x})\) which determines whether the option is exercised in the given state or not. Consequently, errors in the magnitude of \(\hat{T}_k\) are tolerated, while errors in the sign of \(\hat{T}_k\) will lead to degraded performance.
## Results for Bermudan Option Pricing
The 2d example was primarily for illustrative purposes; for a more challenging case we take a 5d asymmetric Max-Call. We keep \(T=3, \Delta t = 1/3, {\cal K}=100, r=0.05\) but adjust the parameters of the log-normal asset dynamics [\[eq:gbm\]](#eq:gbm){reference-type="eqref" reference="eq:gbm"} to be \(\delta_1 = \ldots =\delta_5 = 0.1\) and \(\sigma_i = 0.08 \cdot i\) to have different volatilities. The setting therefore is no longer symmetric and, due to larger variance, the third-through-fifth coordinates are progressively more important. The initial starting point is taken to be \(\mathbf{X}(0) = (70,70,70,70,70)\). We benchmark the results on a fixed test set of \(N' = 25000\) paths of \(\mathbf{X}(t)\), comparing the resulting option price estimates. Our comparators include a global GP model `homGP` and a `hetGP` surrogate. Note that the above are not pre-scaled and take inputs as-is. We train the surrogates with an experimental design sampled from the log-normal density of \(\mathbf{X}(t)\), with \(\bar{N}=2000\) unique inputs and \(a=10\) replicates each. In this case study we found that a relatively small number of inducing points \(m=30\) works well along with a neighborhood size of \(n=50\). We also left out various embellishments that can further fine-tune each surrogate, aiming for a level playing field. Figure [\[tbl:bermudan5d\]](#tbl:bermudan5d){reference-type="ref" reference="tbl:bermudan5d"} presents the results, with the average compute times (in minutes) below each box plot. The reported compute time is for the entire option pricing problem, i.e., across \(K=9\) surrogates. The reference option price is \(11.83\) (black dashed line). With a design of \(\bar{N}=2000\), all three approaches yield similar results that are close to optimal, but the compute times vary substantially. LIGP is more than four times faster than `homGP` and an order of magnitude faster than `hetGP`. In the problem context, methods that take more than a handful of minutes are impractical, and these major speedup gains greatly outweigh the small accuracy loss with LIGP. Not shown in Figure [\[tbl:bermudan5d\]](#tbl:bermudan5d){reference-type="ref" reference="tbl:bermudan5d"} are the results from `laGP`. Using LAGP.ALC with \(n=50\) is prohibitively slow in this example (over an hour, like `hetGP`), while using LAGP.NN yields very poor prices (below 11.5). Thus, `laGP` fails either in terms of speed or in terms of accuracy and is not a viable alternative. Exploration of the input space tends to be the main driver of accuracy, so we consider upgrading to \(\bar{N}=4000\) and \(a=5\) replicates. This keeps the same budget for total number of simulations but is more exploratory. The new choice has minimal impact on the compute time of LIGP that is driven by \(\bar{N} \times a\) and is to zeroth-order independent of \(\bar{N}\). Indeed, the average compute times between \(4000 \times 5\) (6.11 mins) and \(2000 \times 10\) (6.26) are within 3% of each other. On the other hand, \(\bar{N} \gg 10^3\) is too large for a global GP: `homGP` takes nearly three hours with \(\bar{N}=4000\) and `hetGP` is not considered as it takes even longer. Running times that exceed an hour (shown in red in Figure [\[tbl:bermudan5d\]](#tbl:bermudan5d){reference-type="ref" reference="tbl:bermudan5d"}) are prohibitive. A take-away here is that with a global approach one is restricted in the choice of \(\bar{N}\), while LIGP has additional flexibility to survey the exploration-exploitation tradeoff. In this particular setting, having more unique inputs does not yield substantial performance improvement, but in related contexts it is a crucial factor for developing an accurate exercise strategy.
# Discussion {#sec:discussion}
Modern computing resources and advances in MC and agent-based modeling strategies have combined of late to produce large simulation campaigns for complex phenomena. The ability to adequately model such large data sets is limited, especially when the simulator is heteroskedastic. Methods for modeling large data sets exist, such as sparse matrix approximation and divide-and-conquer, but are easily fooled into interpreting noise as signal and provide inadequate UQ in the presence of input-dependent noise. Local approximations (LAGP) on local subsets of data (neighborhoods) perform well on predictive accuracy for deterministic data, because small local subsets yield a substantial computational advantage over competing methods. Scaling up to larger neighborhoods in the face of common replication-based design strategies for separating signal from noise is problematic in that context; larger neighborhoods wipe out that computational advantage. LIGP frees up resources through inducing points, and thus allow for increases in neighborhood size without severe computational bottlenecks, but is developed solely for deterministic data. HetGP was designed for heterogeneous noisy simulations, and replicated responses in the experimental design, and consequently makes thrifty use of sufficient statistics and provides excellent UQ. But it is not a local approximation, and so it does not scale well to the largest simulation campaigns. Here we provide upgrades to the LIGP framework, essentially by combining these three modern methods. We first redefine LIGP neighborhoods based on the number of unique design locations. Applying the common Woodbury identities allows us to take advantage of replication. This provides calculations on the order of the number of inducing points and unique locations. Consequently LIGP may entertain much larger neighborhoods compared to LAGP and the original version of LIGP. With these larger neighborhoods, we present evidence that LIGP is able to better, and more quickly, separate signal from noise. The promising results in this paper provide direction for further areas of study. While current work relies on NN to build local neighborhoods, variance-based alternatives (e.g., `alcray` in `laGP`, ) may enhance modeling the mean surface. Extending the kernel support to other families such as Matérn may also prove useful. Our empirical work relied on default choices of \((m,n)\) without much exploration. A better understanding of these hyperparameters (possibly through Bayesian optimization of out-of-sample RMSE or score) could yield better defaults for simple, immediate implementation. It is worth reinforcing every aspect of our local GP fits is unique to each local neighborhood, providing a discontinuous noise (and mean) surface. This is in contrast to global methods (e.g., HomGP, HetGP). We find that the computational and other model fidelity advantages of such a local approach outweigh the superficial disadvantages of a sometimes "choppy" predictive surface. That said, local-global approach where noise information is shared across neighborhoods could help smooth the noise across the input space. | {'timestamp': '2022-06-01T02:03:47', 'yymm': '2109', 'arxiv_id': '2109.05324', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05324'} |
null | null |
# Introduction[\[sec:in\]]{#sec:in label="sec:in"}
Fock space dualities first appeared in physics in the context of the exploration of the new quantum mechanics in the late 1920s. Wigner and von Neumann showed that, by total antisymmetry, the wave function \[\phi((m_{l1},m_{s1}), \dots, (m_{ln_\text{el}},m_{sn_\text{el}}))\] of \(n_\text{el}\) electrons in an atomic shell with azimutal quantum number \(l\), where \(m_l\) and \(m_s\) are the orbital and spin magnetic quantum numbers, can be expanded on terms of the form \[\label{eq:WvN} \sum_\nu \chi^\lambda_\nu (m_{l1}, \dots, m_{ln_\text{el}}) \psi^\lambda_\nu (m_{s1}, \dots, m_{sn_\text{el}}).\] Here the functions \(\chi^\lambda_\nu\) and \(\psi^\lambda_\nu\) form conjugate bases for conjugate irreducible representations (irreps) of the groups of permutations of their arguments. That is, the matrix representing a given permutation \(s\) in one irrep and the inverse transpose of that representing it in the other one are identical when \(s\) is even and differ by a factor \(-1\) when \(s\) is odd. The equivalence classes of these irreps are described by conjugate Young diagrams \(\lambda\) and \(\tilde\lambda\) like the following pair, where \(\lambda=(2,2,1,1,1)\) in terms of row lengths.
\(\lambda=\) \(\tilde\lambda=\)
The Young diagram \(\lambda\) can have no more than \(2l+1\) rows and no more than 2 columns. Its area equals \(n_\text{el}\). By the correspondence between symmetry and equivalence class of irrep of the general linear group \(\gGL(n)\) of non-singular linear transformations of an \(n\)-dimensional vector space over the complex numbers, or equivalently, its Lie algebra \(\lgl(n)\), the functions [\[eq:WvN\]](#eq:WvN){reference-type="eqref" reference="eq:WvN"} carry products of irreps of \(\lgl(2l+1)\), acting on the arguments \(m_l\), and \(\lgl(2)\), acting on the arguments \(m_s\), and the equivalence classes of these irreps are described by the Young diagrams \(\lambda\) and \(\tilde\lambda\). The atomic shell is a Fock space. The general Fock space is associated with \(k\) kinds of particle, where each kind may be a kind of fermion or a kind of boson. The particles inhabit a common single-kind state space of dimension \(d\). I label the particle kinds by letters \(\tau, \upsilon, \dots\) and basic single-kind states by letters \(p, q, \dots\)Ṫhe creation operator of a particle of kind \(\tau\) in the state \(| p \rangle\) is denoted by \(a^\dagger_{p\tau}\) and the corresponding annihilation operator by \(a_{p\tau}\). These operators obey the usual commutation relations, but contrary to quantum mechanical conventions, \(a^\dagger_{p\tau}\) and \(a_{p\tau}\) are *not assumed Hermitian conjugate*. I impose, indeed, no Hermitian inner product on any state space. The of the system thus described is spanned by the states generated from the vacuum by the operators \[1, \quad a^\dagger_{p\tau}, \quad a^\dagger_{p\tau} a^\dagger_{q\upsilon}, \quad \dots\] It has finite dimension in the case of only fermions and infinite dimension in the presence of bosons. In the atomic example, electrons with spins up and down are viewed as different kinds and the single-kind state space is spanned by the states \(| m_l \rangle\). The following is then an alternative formulation of the observations above. Similar decompositions apply to many different Fock spaces, some of which I describe in this paper. A structure like [\[eq:atom\]](#eq:atom){reference-type="eqref" reference="eq:atom"} is called a in the literature.
# Symplectic and orthogonal groups and Lie algebras and their number conserving and number non-conserving realisations on Fock spaces[\[sec:symor\]]{#sec:symor label="sec:symor"}
The symplectic group \(\gSp(n)\) and orthogonal group \(\gO(n)\) are subgroups of \(\gGL(n)\) defined by non-degenerate bilinear forms \(b\). In terms of basic vectors \(| i \rangle\) for the defining vector space of \(\gGL(n)\), their members \(g\) obey \[\sum_{kl} \langle b | k l \rangle \langle k | g | i \rangle \langle l | g | j \rangle = \langle b | i j \rangle.\] When \(b\) is , the group is , when \(b\) is , it is . For a given \(n\), all symplectic groups are isomorphic and all orthogonal groups are isomorphic, independently of \(b\). I denote the Lie algebras of \(\gSp(n)\) and \(\gO(n)\) by \(\lsp(n)\) and \(\lo(n)\). The group \(\gSp(n)\) only exists for even \(n\), the group \(\gO(n)\) for every \(n\). While \(\gSp(n)\) is simply connected, \(\gO(n)\) is not even connected and its maximal connected subgroup \(\gSO(n)\) of index 2 formed by the orthogonal transformations with determinant 1 not simply connected. (The transformations in the coset have determinant \(-1\).) Both \(\gO(n)\) and \(\gSO(n)\) have double covering groups \(\gPin(n)\) and \(\gSpin(n)\), where the latter is simply connected and a subgroup of the former of index 2. The equivalence classes of irreps of \(\lsp(n)\) and \(\lo(n)\) correspond 1--1 to those of the simply connected groups \(\gSp(n)\) and \(\gSpin(n)\). The of \(\gSpin(n)\) are double valued on \(\gSO(n)\) and likewise those of \(\gPin(n)\) on \(\gO(n)\). The equivalence classes of \(\lsp(n)\) and \(\lo(d)\) irreps are described by Young diagrams with at most \(n/2\) rows. The \(\lsp(n)\) Young diagrams are ordinary ones . To describe spin \(\lo(n)\) irreps (which expand to spin irreps of \(\gSpin(n)\)) one also needs Young diagrams with half-integral row lengths like the following with row lengths \(9/5,7/2,3/2\).
Further, for even \(n\), the Young diagrams of maximal depth come in pairs with opposite signs of the "lengths" of their bottom rows like the following pair for \(n=6\) with row lengths \(4,3,\pm2\).
I call the associated equivalence classes of \(\lo(d)\) irreps. For \(n=2\) the edge whence the single row extends must be specified. (The abelian Lie algebra \(\lo(2)\) actually has a continuum of inequivalent 1-dimensional irreps. Those described by the present Young diagrams are the only ones that occur in the theorems below.) The single-kind state space may be taken as defining vector space for \(\lsp(d)\) and \(\lo(d)\). These Lie algebras then have realisations on any Fock space given by \[\label{eq:nc} g \mapsto \sum_{pq\tau} \langle p | g | q \rangle a^\dagger_{p\tau} a_{p\tau}\] for an arbitrary member \(g\) of the Lie algebra. These operators are seen to conserve the number of particles. I reserve the symbols \(\lsp(d)\) and \(\lo(d)\) for these realisations, which I call *number conserving*. For systems of only fermion or only bosons, the Lie algebras \(\lsp(2k)\) and \(\lo(2k)\) have realisations spanned by the operators \[\label{eq:nnc} \sum_p a^\dagger_{p\tau} a_{p\upsilon} \mp \frac d 2, \qquad \sum_{pq} \, \langle p q | b \rangle a^\dagger_{p\tau} a^\dagger_{q\upsilon}, \qquad \sum_{pq} \, \langle b | p q \rangle a_{p\tau} a_{q\upsilon}\] with \(-\) for fermions and \(+\) for bosons. Here the matrix element \(\langle p q | b \rangle\) of the dual bilinear form is defined by \[\sum_r \langle b | p r \rangle \langle q r | b \rangle = \delta_{pq}.\] The operators in the last two sets in [\[eq:nnc\]](#eq:nnc){reference-type="eqref" reference="eq:nnc"} evidently do not conserve particle number. (The sets are empty for \(k=1\) in the cases of fermions and \(\lo(2k)\) and bosons and \(\lsp(2k)\).) I reserve the symbols \(\lsp(2k)\) and \(\lo(2k)\) for these realisations, which I call *number non-conserving*. The sets of operators [\[eq:nc\]](#eq:nc){reference-type="eqref" reference="eq:nc"} and [\[eq:nnc\]](#eq:nnc){reference-type="eqref" reference="eq:nnc"} commute.
# Helmers's theorem and its orthogonal analogon[\[sec:Hel\]]{#sec:Hel label="sec:Hel"}
In 1961, in the wake of the nuclear BCS theory, Helmers proved the following .
Much later, in 2019, i obtained the following orthogonal analogon of Helmers's theorem .
The proofs of both theorems are based on comparison of the characters on both sides of the equations [\[eq:sp-sp\]](#eq:sp-sp){reference-type="eqref" reference="eq:sp-sp"} and [\[eq:o-o\]](#eq:o-o){reference-type="eqref" reference="eq:o-o"}. Several special cases of these theorems have applications in nuclear physics. Thus for \(k=1\), the number non-conserving Lie algebra \(\lsp(2k)\) is closely related to Kerman's algebra , usually described as an \(\lsu(2)\) algebra. The latter is a Lie algebra over the reals whose complexification is isomorphic to \(\lsp(2)\). The \(\lsp(d)\)--\(\lsp(2)\) duality connects the quantum numbers of quasispin and as exploited extensive in the work of Talmi, in particular . For \(k=2\), the Lie algebra \(\lsp(2k)=\lsp(4)\) is identical to the Lie algebra \(\lo(5)\) introduced by Flowers and Szpikowski to describe systems with both neutrons and protons . For \(k=4\), corresponding to the 4-dimensional space of the nucleonic spin and isospin, \(\lo(2k)\) is the Lie algebra \(\lo(8)\) advanced by these authors as a "quasispin algebra for \(LS\) coupling" . Both Lie algebras \(\lo(5)\) and \(\lo(8)\) attracted attention around the turn of millennium in discussions of pairing in nuclei including isospin \(T=0\) paring.
# Howe's theorem[\[sec:Howe\]]{#sec:Howe label="sec:Howe"}
All dualities mentioned so far are special cases of a very general theorem proved in mathematics by Howe in the 1970s . Presenting his theorem requires more definitions. The following formalism is not that of Howe but a "physicists version" based on second quantisation . The Fock space is the general one with possibly both fermion and bosons. Let \(G\) be a "classical" group , \(\gGL(d)\), \(\gSp(d)\) or \(\gO(d)\), with the single-kind state space as the defining vector space. It has a realisation on the Fock space, which I denote by the same symbol as the abstact group. I denote by \(g\) an arbitrary member of \(G\) and by \(\gamma\) its realisation on the Fock space. For a given \(\tau\), the realisation may be either *cogredient*, that is, \[\gamma a^\dagger_{p\tau} = \left( \sum_q \langle q | g | p \rangle a^\dagger_{q\tau} \right) \gamma, \quad \gamma a_{p\tau} = \left( \sum_q \langle p | g^{-1} | q \rangle a_{q\tau} \right) \gamma.\] for every \(g\), or *contragredient*, that is, \[\gamma a^\dagger_{p\tau} = \left( \sum_q \langle p | g^{-1} | q \rangle a^\dagger_{q\tau} \right) \gamma, \quad \gamma a_{p\tau} = \left( \sum_q \langle q | g | p \rangle a_{q\tau} \right) \gamma.\] for every \(g\). In the cases of \(\gSp(d)\) and \(\gO(d)\), co-and contragredient actions are equivalent because the bilinear form \(b\) provides a similarity transformations between the matrices \(\langle p | g | q \rangle\) and \(\langle q | g^{-1} | p \rangle\). Acting on the operators \(a^\dagger_{p\tau}\) and \(a_{p\tau}\), I define a bracket \(|\cdot,\cdot|\) which equals the anticommutator \(\{\cdot,\cdot\}\) when both operands are fermion operators, otherwise the commutator \([\cdot,\cdot]\), so that their commutation relations can be written \[|a^\dagger_{p\tau},a^\dagger_{q\upsilon}| = |a_{p\tau},a_{q\upsilon}| = 0, \quad |a_{p\tau},a^\dagger_{q\upsilon}| = \delta_{p\tau,q\upsilon}.\] I define another bracket \(]\cdot,\cdot[\) which is the complete opposite, being equal to the commutator when both operands are fermion operators and otherwise the anticommutator. Both brackets can be extended to the span \(\aA\) of the set of creation and annihilation operators. The bracket \(|\cdot,\cdot|\) can be extended further to the set \[\bar \lh = \text{span} \, \{ \,a b \; | \; a,b \in \aA \, \},\] which, by the commutation relations, includes the numbers. By definition, \(|ab,cd|=[ab,cd]\) when either both \(a\) and \(b\) or both \(c\) and \(d\) are fermion operators or both of them are boson operators, and \(|ab,cd|=\{ab,cd\}\) when both \(ab\) and \(cd\) are products of one boson operator and one fermion operator. One can check that this defines \(|\cdot,\cdot|\) unambiguously as a bilinear product on \(\bar \lh\) and that \(\bar \lh\) is closed under the action of \(|\cdot,\cdot|\). In particular \(|h_1,h_2|=[h_1,h_2]=0\) when any one of \(h_1\) and \(h_2\) is a number. The set \(\bar \lh\) equipped with the bilinear product \(|\cdot,\cdot|\) forms a so-called *Lie superalgebra*. When only fermions or only bosons are present, it becomes an ordinary Lie algebra. The set \[\lh = \text{span} \{ \, ]a, b[ \; | \; a,b \in \aA \, \}\] can be shown to be a subalgebra of \(\bar \lh\). Also the pointwise \(G\) invariant subset of \(\lh\), that is, \[\lh^G = \{ \, h \in \lh \; | \; \gamma h = h \gamma \quad \forall g \in G \, \}\] is a subalgebra because, as a subgroup of \(\gGL(d)\), the group \(G\) preserves the bracket \(|\cdot,\cdot|\) on \(\aA\). (The members of \(\gGL(d)\) just change the basis for the single-kind state space.) Explicitly,
1. \(\lh^{\gGL(d)}\) is spanned by the operators \[\begin{gathered}
\sum_p \, ]a^\dagger_{p\tau},a_{p\upsilon}[, \quad (\tau,\upsilon) \in K \times K \cup\bar K \times\bar K, \\ \sum_p \, ]a^\dagger_{p\tau},a^\dagger_{p\upsilon}[, \quad \sum_p \, ]a_{p\tau},a_{p\upsilon}[, \quad (\tau,\upsilon) \in K \times\bar K, \end{gathered}\] where \(K\) and \(\bar K\) are the sets of \(\tau\) with co-and contragredient actions of \(\gGL(d)\).
2. \(\lh^{\gSp(d)}\) and \(\lh^{\gO(d)}\) are spanned by the operators \[\sum_p \, ]a^\dagger_{p\tau},a_{p\upsilon}[, \quad \sum_{pq} \, \langle p q | b \rangle \, ]a^\dagger_{p\tau},a^\dagger_{q\upsilon}[ \,, \quad \sum_{pq} \, \langle b | p q \rangle \, ]a_{q\tau},a_{p\upsilon}[.\]
Howe's theorem now reads: Howe's proof is based in the so-called of the classical groups, which states that the algebra of their invariants is generated by the quadratic invariants. Special cases include fermion and boson \(\gGL(d)\)--\(\lgl(k)\) dualities, fermion \(\gSp(d)\)--\(\lsp(2k)\) and \(\gO(d)\)--\(\lo(2k)\) dualities and boson \(\gSp(d)\)--\(\lo(2k)\) and \(\gO(d)\)--\(\lsp(2k)\) dualities. The boson \(\lo(2k)\) and \(\lsp(2k)\) irreps have infinite dimensions. In particular, the Lie algebra of the \(\gO(d)\)--\(\lsp(2k)\) duality is known in nuclear physics in the case \(d=A\) and \(k=3\), where \(A\) is the mass number, as the complexification of the Lie algebra "\(\gSp(3,\mathbb R)\)" suggested by Rosensteel and Rowe to model nuclear collective motion. Since we are in a boson environment, its irreps are infinite-dimensional. By Howe's theorem, they can be labelled by the known equivalence classes of finite-dimensional irreps of \(\gO(A)\).
# Relation to the dualities with pairs of Lie algebras[\[sec:rel\]]{#sec:rel label="sec:rel"}
Unlike the dualities in Sections [\[sec:in\]](#sec:in){reference-type="ref" reference="sec:in"} and [\[sec:Hel\]](#sec:Hel){reference-type="ref" reference="sec:Hel"}, which relate a couple of Lie algebras, Howe duality relates a group and a Lie (super-)algebra. Howe's theorem does not specify the relation of the equivalence classes of the irreps carried by \(\textup{X}_\lambda\) and \(\Psi_\lambda\). This matter is addressed in later papers by Howe , Rowe, Repka and Carvalho and me . It is fairly easy to show that in the special cases with \(\gGL(d)\) and \(\gSp(d)\), the distinction between the group and its Lie algebra does not matter. (The case of \(\gGL(d)\) is the most complicated one; see .) The case of \(\gO(d)\) is more involved due to the more complicated topology of this group. The equivalence classes of \(\gO(n)\) irreps were identified by Weyl . They are described by ordinary Young diagrams subject to the constraint that no pair of different columns have depths whose sum exceeds \(n\). Rowe, Repka and Carvalho obtained the relation between the equivalence classes of the \(\gO(d)\) and \(\lo(2k)\) irreps in the Howe \(\gO(d)\)--\(\lo(2k)\) duality from an analysis of highest weight states . In 2020, I derived it from the \(\lo(d)\)--\(\lo(2k)\) duality theorem in Section [\[sec:Hel\]](#sec:Hel){reference-type="ref" reference="sec:Hel"} and described it diagrammatically as follows .
The symmetric way in which \(\lo(d)\) and \(\lo(2k)\) enter the \(\lo(d)\)--\(\lo(2k)\) duality theorem suggests that one might similarly derive from this theorem a duality between \(\lo(d)\) and a group. I recently showed that this is indeed true and that the group is \(\gPin(2k)\) . I first constructed a realisation of \(\gPin(2k)\) on the fermion Fock space from the observation that \(\gPin(2k)\) is realised within the Clifford algebra \(\aCl(k)\) , which is realised, in turn, by the algebra generated by the operators \(a^\dagger_{1\tau}\) and \(a_{1\tau}\). More precisely, \(\gPin(2k)\) can be identified with the set of products of linear combinations \(s = \sum_\tau ( \alpha_\tau a^\dagger_{1\tau} + \beta_\tau a_{1\tau})\) obeying \(s^2=-1\) and \(\gSpin(2k)\) with the set of products of an even number of factors of this form . A single \(s\) thus connects the subgroup \(\gSpin(2k)\) and its coset. In particular, in my realisation on the fermion Fock space, \(s = a^\dagger_{11}-a_{11}\) maps to a "partial particle-hole conjugation" \(\sigma\) obeying \[\begin{gathered}
\label{eq:siga} \sigma a_{p1} = (-)^d a^\dagger_{p1} \sigma, \quad \sigma a^\dagger_{p1} = (-)^d a_{p1} \sigma, \\ \sigma a_{p\tau} = (-)^d a_{p\tau} \sigma, \quad \sigma a^\dagger_{p\tau} = (-)^d a^\dagger_{p\tau} \sigma, \quad \tau > 1. \end{gathered}\] One can *define* a set of generalised Young diagrams by the following rules. (i) The rows have either integral or half-integral, positive row lengths, which decrease weakly from top to bottom. (ii) If the row lengths are integral, no pair of different columns have depths whose sum exceeds \(2k\). (iii) If the row lengths are half-integral, the Young diagram has exactly \(k\) rows. One can assign to each such Young diagram properties of a \(\gPin(2k)\) irrep, which my limited space does not allow me to specify. (See ref. for details.) Essentially, the Young diagrams with integral row lengths describe \(\gPin(2k)\) irreps which factor through \(\gO(2k)\), yielding an \(\gO(2k)\) irrep with the same diagram, while those with half-integral row lengths describe \(\gPin(2k)\) irreps which split upon restriction into a pair of mirror \(\lo(2k)\) irreps, one of which has the same Young diagram. My constructed realisation of \(\gPin(2k)\) on the fermion Fock space being denoted by the same symbol, I then proved the following.
# Concluding remarks[\[sec:con\]]{#sec:con label="sec:con"}
The triple of the \(\lo(d)\)--\(\lo(2k)\), \(\gO(d)\)--\(\lo(2k)\) and \(\lo(d)\)--\(\gPin(2k)\) dualities seem to present a nice, unified picture with a high degree of symmetry between the number conserving and the number non-conserving groups or Lie algebras. The only asymmetry is \(\gO(d)\) versus \(\gPin(2k)\), or equivalently, the presence of only non-spin irreps of \(\lo(d)\). This asymmetry is related to \(2k\) being always even while both even and odd values are allowed for \(d\). My proofs of the \(\gO(d)\)--\(\lo(2k)\) and \(\lo(d)\)--\(\gPin(2k)\) dualities is based on the \(\lo(d)\)--\(\lo(2k)\) duality, whose proof is based, in turn, on a comparison of characters. In particular the proof of the \(\lo(d)\)--\(\gPin(2k)\) duality does not require a \(\gPin(n)\) first main theorem, and in fact, such a theorem is not known . Application of the method is restricted, though, to the fermion case because it employs Weyl's formula for the character of a finite-dimensional irrep of a semi-simple Lie algebra over the complex numbers. | {'timestamp': '2021-09-14T02:15:30', 'yymm': '2109', 'arxiv_id': '2109.05335', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05335'} |
# Introduction {#sec:introduction}
Recently, pre-trained deep learning models have proven their effectiveness in a variety of tasks. One consequence of this is that the linguistic resources received less attention. In the digital age, the amount of data generated per second is enough to train a model capable of abstracting many different patterns. With a sufficiently large amount of data, these models are shown to be able to abstract the linguistic features on their own. Even so, training the model unsupervised using data on the internet can generate noisy, biased models and make deep learning black boxes to humans. Therefore, we want to find a balanced solution between those two extremes. Pre-trained models derive their power from the attention mechanism especially multi-head self-attention. This mechanism mimics human observation, i.e. only cares about important aspects of the data that are relevant to the decision that needs to be made. The model using attention mechanisms are often learned from data through unsupervised tasks. This helps to increase the performance of tasks without requiring much human intervention. Even so, automation without explanation can carry risks when it is widely adopted. Training models using only data can produce highly accurate but still fallacious models. Table [1](#tab:examples){reference-type="ref" reference="tab:examples"} shows a small dataset where if the model is based on the occurrence of the word \"is\", it can correctly classify 100% of the sentiment labels of the samples. Linguistically, \"is\" is a functional word and should not be used to make a decision in this case. This is a contrived example, but the same thing can happen for a huge data set that humans are not capable of verifying. For that reason, before this work, there are proposals trying to force deep learning models to follow the knowledge of linguistic resources. We observe that such coercion could reduce the flexibility of the model and overall efficiency in exceptional cases.
Finding a neutral solution, we investigate the possibility of injecting some linguistic knowledge for pretrained models to refer to without completely depending on them. With this approach, the model can both use linguistic knowledge for the majority of cases and observe actual cases appearing in the data to adjust its understanding. In addition, this approach guides the model to use the features that are accepted by humans, so that their decision-making is explainable by humans. The challenge of this approach is to find a balance between the information learned in the pretraining and the injected knowledge. We propose an unprecedented approach to incorporating linguistic knowledge into transformers architecture. This approach has 3 advantages compared to other methods. First, this is a knowledge injection method that can take the edge of many different types of annotated resources in NLP. Second, the injected information is not rigid but only a reference source for the model to reach its final conclusion. Third, training and storing the appended component are extremely efficient, helping to solve environmental and financial problems. Although we only experiment with dependency structure as the knowledge source in the scope of this paper, the approach is general to all linguistic knowledge which expresses relationships between internal elements of a sentence, is friendly with transformer architecture, and is extendable. We believe this approach will create a new research direction in injecting knowledge into the transformer architecture. Our paper includes the following main contributions as follows:
- We first propose a novel framework to pretrain attention heads separately using the linguistic information without pretrain the whole model again.
- Next, we introduce HYDRA, a knowledge-injected component to the transformer model, which is pretrained with dependency structure relations.
- Finally, from the experimental result, we draw conclusions and future directions of this research.
In the remainder of this paper, we present related work in the attention mechanism and knowledge injection methods in Section [2](#sec:background){reference-type="ref" reference="sec:background"}, the detail of our method in Section [3](#sec:method){reference-type="ref" reference="sec:method"}. Next, we investigate the effectiveness of the proposed method in comparison with the baseline in Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"}. Finally, the conclusion and future works are presented in Section [5](#sec:conclusions){reference-type="ref" reference="sec:conclusions"}.
# Related Work {#sec:background}
## Attention Mechanism
Transformer has become the architecture that drastically changes the way deep learning is designed and used. With previous models such as LSTM or CNN, all the information of the inputs would be compressed into a single representation. The human need in NLP is to solve longer and longer sentences. With long inputs, these models run into problems, their representations are not able to store all the information of the input. Instead of storing a loop of information in the architecture like the LSTM, the transformer uses the attention mechanism to perceive constraints in the inputs. Attention can be described as aggregating a query and key-value pairs to the representation. The output (attention vector) is calculated as the weighted sum of the values, where the weight assigned to each value is computed by the query with the corresponding key. In the paper, the authors recommend of a combination of multiple aspects to the final representations, so-called multi-head attention. This mechanism allows the model to view data with different aspects to make decisions.
## Knowledge Injection
Knowldge injection is the method of providing deep learning model with external knowledge. In their paper, Nguyen and colleagues use the similarity of the distribution as additional knowledge for the learning process. The knowledge was injected into the neural network (LSTM) to guide the model to make better predictions. The author also conduct the experiment to extract the knowledge from the data and use it to guide the model in another task. However, with the advent of transformers, LSTM gradually became obsolete. Pretrained transformer-based language models such as BERT have demonstrated their effectiveness in many NLP tasks. However, these bulky models also reveal weaknesses through unsupervised learning of data. As a result, using additional knowledge is a promising way to improve the models. Recent works discover that the performance of transformer models could be improved by incorporating either semantic information or syntactic information into the transformer models. To encode syntactic information into a transformer model, the authors consider using a syntactic representation to incorporate it into the attention mechanism so-called Syntax-guided network(SG-NET). The key component in SG-NET is the SDOI-SAN, which is a combination of syntactic dependency of interest (SDOI) and self-attention network (SAN). SG-NET used both SDOI-SAN and SAN of the original transformer model for better linguistics-inspired representation. This is a matrix \(n \times n\) with \(n\) as the number of words in the input sentence. An element in this metric is assigned a value of 1 if there is a dependency link of the words in the corresponding row and column, otherwise, it is assigned to 0. With the same idea of injecting knowledge to the model, Zhang et al. proposes an architecture with combined semantic information and attention mechanism, so-called Semantics-guided neural network (SGN). The authors assign semantic labels for every token, encode them and integrate this kind of knowledge to the pretrained model using a convolutional neural network. We observe that in terms of implementation, this design is not friendly with the existing architecture of transformers and is hard to extend.
# Method {#sec:method}
Observing the limitations of the previous methods, in this paper, we propose a novel partial pretraining method on the Transformer architecture. Attention heads are essentially abstractions of data from different perspectives formed during the pretraining process. The information in the heads interacts with each other to make the final decision. So, we want to verify the possibility of pretraining the heads seperately with linguistic information and then incorporate them into the general architecture as an ordinary component. The overall architecture of the proposed framework is shown in Figure [\[fig:framework\]](#fig:framework){reference-type="ref" reference="fig:framework"}. Based on the transformer's architecture, we add one more transformer layer at the end of the architecture containing heads, which are pretrained with linguistic knowledge. The model can choose to use the information of these heads or ignore it with the residual connections. Therefore, we can use linguistic knowledge as a reference source for the model without forcing it to accept such knowledge rigidly. Since the idea is derived with a dependency structure, we give this architecture an inspiring name, *HYDRA*, which stands for *Hyper Dependency Representation Attentions*.
## Pretraining HYDRA Heads
Let the input be \(X = \{x_1, x_2,..., x_n\}\) with \(n\) be the sequence length, we obtain an \(n \times n\) matrix \(M^*\) containing the linguistic relationship of the words as introduced by Zhang et al. . After passing the input through the transformer layers, we get hidden state \(H_{l}\) at the last layer with \(l\) be the number of layers of the original transformer body. We initiate and append \(l+1^{th}\) layer containing HYDRA heads and pass \(H_{l}\) to this layer. With \(W_q\), \(W_k\) are learnable parameters in the layer \({l+1}^{th}\), we calculate query and key vectors as in Equations [\[eq:qvec\]](#eq:qvec){reference-type="ref" reference="eq:qvec"} and [\[eq:kvec\]](#eq:kvec){reference-type="ref" reference="eq:kvec"}: \[\begin{aligned}
q_{l+1}^h &= W_q \cdot H_{l} \label{eq:qvec} \\ k_{l+1}^h &= W_k \cdot H_{l} \label{eq:kvec} \end{aligned}\] With \(d_k\) be the dimension of the key matrix \(k_{l+1}^h\), we calculate the attention matrix for each head \(M^h\) following Equation [\[eq:attn_mtx\]](#eq:attn_mtx){reference-type="ref" reference="eq:attn_mtx"}: \[\label{eq:attn_mtx} M^h = \frac{q_{l+1}^h \cdot {k_{l+1}^h}^\intercal}{\sqrt{d_k}}\] We then minimize the element-wise MSE loss (\(\mathcal{L}\)) between \(M^*\) and \(M^h\) for every HYDRA head: \[\begin{aligned}
\mathcal{L} &= \frac{1}{n^2} \sum_{i=0}^{n^2} (M^*_i-M^h_i)^2 \end{aligned}\] where \(i\) \((0 \leq i < n^2 )\) is the index of element in flatten attention matrix SDOI (\(M^*_i\)). The weights of the transformer body are frozen and the weights in \(Q_{l+1}^h\) and \(K_{l+1}^h\) in every HYDRA heads are updated via the backpropagation process. After pretraining, we get the HYDRA heads corresponding to the transformer body coupled to it. These HYDRA heads are saved in the storage and ready to be delivered as a form of deep learning resource. Compared to storing pretrained transformer models, storing HYDRA heads saves much more storage space. A transformer model checkpoint can be measured in gigabytes, while a HYDRA head checkpoint in our experiment is less than 10 megabytes.
## Fine-tuning with HYDRA Heads
Different from approaches that use linguistic knowledge that is rigidly injected into the model, we attach pretrained HYDRA heads to the transformer bodies. These models can refer to but are not limited by the linguistic knowledge learned in the HYDRA heads. At this phase, the weights of the transformer body and HYDRA head are both updated to match the training data. The final model now contains \(l+1\) layers, first \(l\) layers are from the original transformer body and the last layer contains HYDRA heads. We simply pass the \(H_{l+1}\) to a fully connected layer to get the logits for each downstream task. The loss function of each downstream task is defined similarly to the work. We have experiments to observe the performance of the model in Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"}. Intuition in this phase is that by acquiring linguistic knowledge, the system can properly model the problem rather than try to fit into the local minima.
# Experiments {#sec:experiments}
## Experimental Settings
### Pretraining Stage
In the pretrain phase, we collect data from Wikimedia Downloads[^1] using WikiExtractor package . We use the variant of dependency parser described by Honnibal and Johnson , provided by spaCy[^2]. After data processing, we obtain 330,000 samples for training and 50,000 samples for validation. We use BERT as the base model to pretrain the HYDRA heads. We set the maximum sequence length to 512 and only use sentences whose length does not exceed this number to pretrain the HYDRA heads. The goal of this setting is to help the model learn the linguistic structure of complete sentences of the longest possible length. With 1.8 million parameters, in our experiment, it takes only one or two epochs for the HYDRA heads to reach optimal loss on both the training set and the validation set.
### Task Adaptation Stage {#sec:task_adap}
To understand the model's behavior and the method's effectiveness, we run experiments and compare the performance of the HYDRA variants with the corresponding non-HYDRA baselines. The experiments are run on datasets and standard settings of several famous benchmarks in natural language processing which include:
- QNLI , MNLI , RTE : Natural Language Inference and Texture Entailment
- Metric for QNLI and RTE: Accuracy.
- Metric for MNLI: Accuracy for both matched (m) and mismatched (mm) versions.
- QQP and STS-B : Semantically Equivalence Judgment
- Metric for QQP: Accuracy.
- Metric for STS-B: Pearson Spearman Correlation.
- SQuAD and SQuAD 2.0 : Question Answering
- Metric for SQuAD and SQuAD 2.0: Exact Match and Macro-averaged F1 Score.
With these benchmarks, we can investigate in detail the effectiveness of the proposed method. We conduct experiments with both the cased and uncased versions of BERT and average the performance.
## Experimental Result and Discussion
Table [2](#tab:result){reference-type="ref" reference="tab:result"} shows the results of the models on the benchmarks and metrics described in Section [4.1.2](#sec:task_adap){reference-type="ref" reference="sec:task_adap"}. BERT is a very strong baseline, it achieved high performance on all benchmarks. Even so, our model still can slightly improve the results of BERT with the appended HYDRA heads. This result supports the hypothesis that linguistic knowledge pretrained in HYDRA heads can boost the performance of the vanilla model. In addition to the performance improvement, this is a novel approach to inject linguistic knowledge into language models. One notable feature of our improvement is that this new component is lightweight, requires low pretraining computation cost and storage. With this paradigm, we can improve the bulky transformers models without pretraining the whole network again or force them to follow rigid linguistic rules.
# Conclusions {#sec:conclusions}
This paper proposes an architecture-friendly and extensible method to improve the effectiveness of the transformer-based language models by pretraining and appending new knowledge-guided heads to their architecture. We conduct the experiment with BERT as the base model and the dependency information as the external knowledge. Our experiment shows that our lightweight component can help to boost the performance of transformer models and provide a flexible paradigm to partially inject the knowledge into bulky models without pretraining them again. Extending this work, we can analyze the possibility of pretraining this component with different knowledge forms for problems in narrower domains or explore the potential of this approach for other data such as photos or videos.
[^1]: https://dumps.wikimedia.org/
[^2]: https://spacy.io/ | {'timestamp': '2021-09-14T02:15:58', 'yymm': '2109', 'arxiv_id': '2109.05349', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05349'} |
# Introduction
The various statistics have been proposed to describe the phenomena which show power-like distributions. An extension of the Boltzmann-Gibbs statistics is the Tsallis statistics, and the statistics has been applied in various branches of science. The escort average is often adopted to calculate the physical quantities in the Tsallis statistics. The Tsallis statistics has the entropic parameter \(q\), and the statistics approaches the Boltzmann-Gibbs statistic as \(q\) approaches one. The entropic parameter \(q\) is often restricted. The normalizability of the probability requires that \(q\) is less than two. The parameter \(q\) is also restricted because physical quantities are restricted. For example, the energy density should be finite and the number of particles should be positive, and these requirements show that the maximum value of \(q\) is smaller than two. The limitation of \(q\) was also derived by using the conjugate variables theorem. Simple systems have been adopted to study the effects of statistics. The classical gas model was adopted, and it was found that the energy is proportional to the number of particles and the physical temperature in the Tsallis statistics. It is also found that the number of the particles are restricted. The thermodynamic quantities for a classical harmonic oscillator was also calculated in the Tsallis statistics with escort average. The partition function was calculated and the energy was obtained. The calculations of the thermodynamic quantities for the harmonic oscillators are required in the Tsallis statistics. A field is decomposed into harmonic oscillators with different frequencies to calculate physical quantities. The results for the harmonic oscillators with different frequencies in the Tsallis statistics will be helpful to calculate physical quantities in various systems. In this paper, we study the thermodynamic quantities in the system of the \(N\) independent harmonic oscillators in the Tsallis statistics of the entropic parameter \(q\). The range of \(q\) is set between one and two in this study. The escort average is employed to obtain physical values. In Sec. [2](#sec:review-Tsallis){reference-type="ref" reference="sec:review-Tsallis"}, we briefly review the Tsallis statistics. In Sec. [3](#sec:N-oscillators){reference-type="ref" reference="sec:N-oscillators"}, we study the \(N\) independent harmonic oscillators with different frequencies. The self-consistent equation is derived, and the equation is solved approximately. The expression of the energy is obtained with physical temperature. The expressions of Tsallis and Rényi entropies are also obtained. The last section is assigned for conclusion.
# Brief review of the Tsallis statistics {#sec:review-Tsallis}
The Tsallis statistics is based on the Tsallis entropy \(S_q^{(T)}\) with the entropic parameter \(q\). The entropy \(S_q^{(T)}\) is defined by \[\begin{aligned}
S_q^{(T)} = \frac{1-\Tr{\hrho^q}}{q-1}, \end{aligned}\] where \(\hrho\) is the density operator. The density operator \(\hrho\) is obtained by extremizing \(S_q^{(T)}\) under the normalization condition \(\Tr \hrho = 1\) and the energy constraint: \[\begin{aligned}
U = \frac{\Tr{\hrho^q \hH}}{\Tr{\hrho^q}}, \label{eq:constraint:U} \end{aligned}\] where \(U\) is the energy. The right-hand side of Eq. [\[eq:constraint:U\]](#eq:constraint:U){reference-type="eqref" reference="eq:constraint:U"} is the escort average of the Hamiltonian \(\hH\). The density operator \(\hrho\) in the Tsallis statistics with the escort average is obtained:
where \(\beta\) is the inverse temperature. The partition function \(Z\) is related to \(c_q\): \[\begin{aligned}
c_q = Z^{1-q}. \label{c_q:Z} \end{aligned}\] The inverse physical temperature is given by \[\begin{aligned}
\betaph = \beta / c_q. \end{aligned}\] The physical temperature \(\Tph\) is given as \(1/\betaph\). The thermodynamic quantities are calculated with the above density operator for the \(N\) independent harmonic oscillators with different frequencies in the following section.
# The independent harmonic oscillators with different frequencies {#sec:N-oscillators}
## Derivation of self-consistent equation
We attempt to derive the self-consistent equation by calculating \(c_q\) in two ways. One way is the method by using the relation \(c_q = Z^{1-q}\) and the other way is the method by calculating \(c_q = \Tr{\hrho^q}\) directly. We obtain the self-consistent equation by equating these results. We treat the \(N\) independent harmonic oscillators with different frequencies. The Hamiltonian \(\hH\) is \[\begin{aligned}
\hH = \sum_{j=1}^N \omega_j \left( \hn_j + \frac{1}{2} \right), \end{aligned}\] where \(\hat{n}_j\) is the number operator with the subscript \(j\). We treat the above Hamiltonian in the Tsallis statistics of \(1<q<2\): \((2-q)/(q-1)\), \(1/(q-1)\), and \(q/(q-1)\) are positive. We introduce a parameter \(\Eref\) and calculate the partition function \(Z\): \[\begin{aligned}
Z &= \sum_{n_1, \cdots, n_N = 0}^{\infty} \left\{ 1+(q-1) \betaph \Bigg( \frac{\hbar}{2}(\omega_1+\cdots+\omega_N)-U \Bigg) + (q-1) \betaph \Bigg( \hbar \omega_1 n_1 + \cdots + \hbar \omega_N n_N \Bigg) \right\}^{\frac{1}{1-q}} \nonumber \\ & = \left( (q-1)\betaph \Eref \right)^{\frac{1}{1-q}} \sum_{n_1, \cdots, n_N = 0}^{\infty} (\lambda_N + a_1 n_1 + \cdots + a_N n_N)^{\frac{1}{1-q}}, \label{eqn:Z} \end{aligned}\] where
Equation [\[eqn:Z\]](#eqn:Z){reference-type="eqref" reference="eqn:Z"} is represented with Barnes zeta function \(\zeta_{\mathrm{B}}(s, \alpha| \vec{\omega}_N)\) (See Eq. [\[def:Barnes-zeta\]](#def:Barnes-zeta){reference-type="eqref" reference="def:Barnes-zeta"}): \[\begin{aligned}
Z & = \left( (q-1)\betaph \Eref \right)^{\frac{1}{1-q}} \zeta_{\mathrm{B}}\left(1/(q-1), \lambda_N | \vec{a}_N \right) \qquad \vec{a}_N = (a_1, a_2, \cdots, a_N). \label{eqn:Z:Banes} \end{aligned}\] The condition \(s>N\) for the parameters of the Barnes zeta function in the present case is \[\begin{aligned}
\frac{1}{q-1} > N. \end{aligned}\] This means that the number of the oscillators is restricted. We also calculate \(c_q\) directly as \[\begin{aligned}
c_q = \Tr{\hrho^q} = Z^{-q} \left( (q-1)\betaph \Eref \right)^{\frac{q}{1-q}} \zeta_{\mathrm{B}}\left(q/(q-1), \lambda_N | \vec{a}_N \right). \label{cq-from-def} \end{aligned}\] From Eqs. [\[c_q:Z\]](#c_q:Z){reference-type="eqref" reference="c_q:Z"}, [\[eqn:Z:Banes\]](#eqn:Z:Banes){reference-type="eqref" reference="eqn:Z:Banes"}, and [\[cq-from-def\]](#cq-from-def){reference-type="eqref" reference="cq-from-def"}, we have the following self-consistent equation: \[\begin{aligned}
\left( (q-1)\betaph \Eref \right) \zeta_{\mathrm{B}}\left(1/(q-1), \lambda_N | \vec{a}_N \right) = \zeta_{\mathrm{B}}\left(q/(q-1), \lambda_N | \vec{a}_N \right). \label{self-consistent-eq} \end{aligned}\] We attempt to obtain the physical quantities by solving the self-consistent equation in the next subsection.
## Energy and entropies
We attempt to find the expressions of physical quantities in this subsection. For \(\lambda_N \gg 1\), we have the following expressions by using Eq. [\[eqn:Bzeta:approximation\]](#eqn:Bzeta:approximation){reference-type="eqref" reference="eqn:Bzeta:approximation"}.
In [\[Barnes-zeta\]](#Barnes-zeta){reference-type="ref" reference="Barnes-zeta"}, the above approximated expression for the Barnes zeta function is given by using the approximated expression for the Hurwitz zeta function given in [\[hurwitz-zeta\]](#hurwitz-zeta){reference-type="ref" reference="hurwitz-zeta"}. We use these expressions of \(\zeta_{\mathrm{B}}\) to solve Eq. [\[self-consistent-eq\]](#self-consistent-eq){reference-type="eqref" reference="self-consistent-eq"} approximately.
### Expression of the energy
We attempt to calculate the energy \(U\) by solving Eq. [\[self-consistent-eq\]](#self-consistent-eq){reference-type="eqref" reference="self-consistent-eq"}. Substituting Eqs.[\[eqn:approximate1\]](#eqn:approximate1){reference-type="eqref" reference="eqn:approximate1"} and [\[eqn:approximate2\]](#eqn:approximate2){reference-type="eqref" reference="eqn:approximate2"} into Eq. [\[self-consistent-eq\]](#self-consistent-eq){reference-type="eqref" reference="self-consistent-eq"}, we have \[\begin{aligned}
U = \frac{\Tph}{(q-1)} \left( 1-\frac{\displaystyle\prod_{j=0}^{N-1} ((2-q)-j(q-1)) }{ \displaystyle\prod_{j=0}^{N-1} (1-j(q-1)) } \right) + \sum_{i=1}^N \frac{\hbar\omega_i}{2} , \qquad N < \frac{1}{(q-1)}. \label{expression-U} \end{aligned}\] It is noted that Eq. [\[expression-U\]](#expression-U){reference-type="eqref" reference="expression-U"} does not contain \(\Eref\). We obtain easily \[\begin{aligned}
\frac{\displaystyle\prod_{j=0}^{N-1} ((2-q)-j(q-1)) }{ \displaystyle\prod_{j=0}^{N-1} (1-j(q-1)) } = 1-N(q-1). \end{aligned}\] By substituting the above expression, we have the following expression of \(U\): \[\begin{aligned}
U = N \Tph + \frac{1}{2} \sum_{i=1}^N \hbar\omega_i, \qquad N < \frac{1}{(q-1)}. \label{eqn:U} \end{aligned}\] Equation [\[eqn:U\]](#eqn:U){reference-type="eqref" reference="eqn:U"} is the well-known form of the energy \(U\) in the Boltzmann-Gibbs statistics. It is possible to evaluate \(\lambda_N\) by using Eq. [\[eqn:U\]](#eqn:U){reference-type="eqref" reference="eqn:U"}: \[\begin{aligned}
\lambda_N = \frac{1-N (q-1)}{(q-1)\betaph \Eref}. \label{eqn:lambdaN:eval} \end{aligned}\] The numerator of the right-hand side of Eq. [\[eqn:lambdaN:eval\]](#eqn:lambdaN:eval){reference-type="eqref" reference="eqn:lambdaN:eval"} is positive, because \(N (q-1)\) is less than one. Therefore the condition \(\lambda_N \gg 1\) is satisfied for \((q-1) \betaph \Eref \ll 1\): the condition is satisfied at high physical temperature \(\Tph\) and/or for small deviation \((q-1)\).
### Expressions of the entropies
The Tsallis entropy \(S_q^{(T)}\) is represented as \[\begin{aligned}
S_q^{(T)} = \frac{1-c_q}{q-1} = \frac{1-Z^{1-q}}{q-1}. \end{aligned}\] The Rényi entropy \(S_q^{(R)}\) is related to the Tsallis entropy: \[\begin{aligned}
S_q^{(R)} = \frac{1}{1-q} \ln (1+(1-q)S_q^{(T)}). \end{aligned}\] This equation is represented with \(c_q\) as \[\begin{aligned}
S_q^{(R)} = \frac{1}{1-q} \ln c_q = \frac{1}{1-q} \ln e^{(1-q) \ln Z} = \ln Z. \label{eqn:RenyiEntropy} \end{aligned}\] We calculate \(Z\) approximately by using Eq. [\[eqn:approximate1\]](#eqn:approximate1){reference-type="eqref" reference="eqn:approximate1"}. \[\begin{aligned}
Z =\frac{1}{ \left( \displaystyle\prod_{j=0}^{N-1} ((2-q)-j(q-1)) \right) \left( \displaystyle\prod_{j=1}^{N} (\betaph \hbar \omega_j) \right) \left( 1 + (q-1) \betaph \left( \displaystyle\frac{1}{2} \displaystyle\sum_{i=1}^N (\hbar \omega_i)-U \right)\right)^{\frac{1}{q-1}-N} }. \label{eqn:Z:approximation} \end{aligned}\] Substituting Eq. [\[expression-U\]](#expression-U){reference-type="eqref" reference="expression-U"} into Eq. [\[eqn:Z:approximation\]](#eqn:Z:approximation){reference-type="eqref" reference="eqn:Z:approximation"}, we obtain \[\begin{aligned}
Z = \frac{\Bigg( \displaystyle\prod_{j=0}^{N-1} (1-j(q-1)) \Bigg)^{\frac{1}{1-q}-N}} {\Bigg( \displaystyle\prod_{j=1}^{N} (\betaph \hbar \omega_j) \Bigg) \Bigg( \displaystyle\prod_{j=0}^{N-1} ((2-q)-j(q-1)) \Bigg)^{\frac{q}{1-q}-N}}. \label{eqn:Z:approximation:q-1:rep} \end{aligned}\] We find the relation between \(dU\) and \(dS_q^{(R)}\). The Rényi entropy is given by \(\ln Z\). For the fixed \(N\) and \(q\), we have \[\begin{aligned}
dS_q^{(R)} = d\ln Z = N \frac{d\Tph}{\Tph}. \label{eqn:dSR} \end{aligned}\] With Eqs. [\[eqn:U\]](#eqn:U){reference-type="eqref" reference="eqn:U"} and [\[eqn:dSR\]](#eqn:dSR){reference-type="eqref" reference="eqn:dSR"}, we have \[\begin{aligned}
dU = N d\Tph = \Tph dS_q^{(R)}. \end{aligned}\] The \(q\)-dependence of \(S_q^{(R)}\) for small \(q-1\) is obtained by expanding the logarithm of Eq. [\[eqn:Z:approximation:q-1:rep\]](#eqn:Z:approximation:q-1:rep){reference-type="eqref" reference="eqn:Z:approximation:q-1:rep"} with respect to \(q-1\). We have \[\begin{aligned}
S_q^{(R)} = \ln Z = L_N(\Tph) + N + \frac{1}{2} N (q-1) + O((q-1)^2), \label{eqn:lnZ} \end{aligned}\] where \(L_N(\Tph)\) is defined by \[\begin{aligned}
L_N(\Tph) := \sum_{j=1}^{N} \ln \Big(\frac{\Tph}{\hbar \omega_j}\Big). \end{aligned}\] The same equation can be obtained by substituting Eq. [\[eqn:U\]](#eqn:U){reference-type="eqref" reference="eqn:U"} into Eq. [\[eqn:Z:approximation\]](#eqn:Z:approximation){reference-type="eqref" reference="eqn:Z:approximation"}. We remember that \(N(q-1)\) is less than one. We obtain the ratio of \(S_q^{(T)}\) to \(S_q^{(R)}\). Hereafter we omit the argument \(\Tph\) of \(L_n\) for simplicity. For \(L_N \gg N\), we have
The ratio \(S_q^{(T)}/S_q^{(R)}\) is \[\begin{aligned}
\frac{S_q^{(T)}}{S_q^{(R)}} \sim \frac{1-e^{-(q-1) L_N}}{(q-1)L_N}. \end{aligned}\] The ratio \(S_q^{(T)}/S_q^{(R)}\) is approximately \(1/((q-1) L_N)\) at sufficiently high physical temperature which satisfies \((q-1) L_N \gg 1\). This ratio is \(1-(q-1) L_N/ 2\) for \((q-1) L_N \ll 1\), though the condition \(L_N \gg N\) is required to obtain the expressions, Eqs. [\[eqn:ST-highTph\]](#eqn:ST-highTph){reference-type="eqref" reference="eqn:ST-highTph"} and [\[eqn:SR-highTph\]](#eqn:SR-highTph){reference-type="eqref" reference="eqn:SR-highTph"}.
# Conclusions {#sec:conclusion}
We studied the thermodynamic quantities in the system of the \(N\) independent harmonic oscillators with different frequencies \(\omega_j\) in the Tsallis statistics of the entropic parameter \(q\) (\(1<q<2\)). The number of the oscillators \(N\) was fixed and the escort average was adopted in this study. We derived the self-consistent equation, and the expressions of physical quantities with the physical temperature were obtained. We obtained the partition function \(Z\), the energy \(U\), the Rényi entropy \(S_q^{(R)}\), and the Tsallis entropy \(S_q^{(T)}\) by solving the self-consistent equation approximately at high physical temperature \(\Tph\) and/or for small deviation \(q-1\). It was found from the condition for the parameters of the Barnes zeta function that the number of harmonics oscillators \(N\) is less than \(1/(q-1)\). The restriction of the number of the harmonic oscillators exists, as the restriction was previously given for the classical gas. As expected, the supremum \(1/(q-1)\) goes to infinity when \(q\) approaches one. The energy \(U\) is \(q\)-independent at high physical temperature when the physical temperature is adopted. The energy is proportional to the number of harmonic oscillators \(N\) and the physical temperature \(\Tph\) at high physical temperature when the vacuum term is ignored: the expression of the energy is the well known expression, \({U = N \Tph + \sum_j \hbar\omega_j/2}\). The Rényi entropy \(S_q^{(R)}\) is the sum of the values for the independent harmonic oscillators at high physical temperature. The Rényi entropy with the same frequency, \(\omega \equiv \omega_1 = \cdots = \omega_N\), is given by \(N \ln(\Tph/(\hbar \omega))\) which is well-known expression for the \(N\) independent harmonic oscillators with the same frequency. The Tsallis entropy \(S_q^{(T)}\) was obtained through the Rényi entropy. The variation for the Rényi entropy is simply given as \(dS_q^{(R)} = N d\Tph\) for the fixed \(N\), and the well-known relation between \(dU\) and \(dS_q^{(R)}\) is also obtained: \(dU = \Tph dS_q^{(R)}\). The physical temperature dependence of the Tsallis entropy is different from that of the Rényi entropy. The Rényi entropy contains the term that is \(\sum_{j=1}^N \ln(\Tph/(\hbar \omega_j))\). Therefore, the Rényi entropy increases with the physical temperature, and is unbounded from the above. In contrast, the Tsallis entropy increases with the physical temperature, and is bounded from the above. The ratio of the Tsallis entropy to the Rényi entropy, \(S_q^{(T)}/S_q^{(R)}\), is small at high physical temperature. The difference between the Tsallis and Rényi entropies is large at high physical temperature. The system of the independent harmonic oscillators with different frequencies is basic, and the results in this study will give the insight on other physical systems. The author believes that the present study will be helpful for the reader to study the system represented with oscillators in unconventional statistics such as the Tsallis statistics. | {'timestamp': '2021-09-14T02:19:46', 'yymm': '2109', 'arxiv_id': '2109.05442', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05442'} |
null | null |
null | null |
# Introduction
Systems displaying glassy dynamics are prevalent in nature ranging from granular matter to colloids to biological applications. For example, experiments demonstrate that bacterial cytoplasms become glassy in the case of a suppressed metabolism . Recently, the importance of glassy physics has also been revealed in numerous studies on active particles . Furthermore, significant progress on the slow structural relaxation has been achieved by using machine learning techniques and novel optimized algorithms in computer simulations . Yet, the nature of the glass transition is still heavily debated and remains one of the grand challenges of condensed matter physics . The mode-coupling theory of the glass transition (MCT) provides an entirely first-principle description to elucidate this complex behavior . Other approaches are dynamical mean-field theory and self-consistent generalized Langevin dynamics . Only recently, mode-coupling-like approaches have been successfully extended to systems of active Brownian particles and self-propelled particles . Indisputably, confined systems are of particular interest due to their presence in nature and industrial applications. Since local structure plays a predominant role for the slowing down of the dynamics upon approaching the glass transition, confinement strongly affects both structure and dynamics. This complicated interplay of confinement with the cage effect has been extensively investigated within the slit geometry in both experiments and hard-sphere simulations . MCT has been successfully extended to these inhomogeneous confined liquids using symmetry-adapted modes . Other inhomogeneous MCT approaches incorporate (infinitesimal) local external fields or periodic potential landscapes . Although its relative simplicity, which makes it one of the most widely studied confined systems, the slit geometry already incorporates a non-trivial interplay between local packing (caging) and the inhomogeneous density profile induced by the walls (layering). However, in a slit geometry it is not possible to unravel the effects of confinement and those simply due to layering. To circumvent this problem, quasi-confined liquids are considered , where periodic boundary conditions in the short confining direction are introduced. In this set-up, the liquid becomes translationally invariant both in the unconfined as well as in the confined direction, with a uniform density profile. Yet, there is still an impact from the confinement on the particles once the characteristic confinement length becomes comparable to the particle size. The goal of the present work is to advance the analysis of quasi-confined liquids within the framework of MCT and with event-driven molecular dynamics simulations. The collective behavior has been discussed in detail recently . Even in absence of layering a reentrance behavior is observed in the non-equilibrium state diagram, motivating us to further investigate these liquids, in particular with regard to the self-dynamics. In principle, the tagged-particle motion is directly accessible in experiments using neutron scattering, dynamic light scattering or single particle tracking. Due to enhanced statistics it is also favorable to extract self-correlation functions from computer simulations, which is of particular interest for this work. The tagged-particle MCT equations for the general class of layered fluids have been derived earlier for Newtonian dynamics and have been extended to Brownian dynamics more recently . As for the collective dynamics, the only required input to these MCT equations are the static quantities of quasi-confined liquids . If the self-intermediate scattering function (SISF) is known, all moments of the displacement such as the mean-square displacement (MSD) can be extracted. The long-time behavior of the MSD provides the diffusion coefficient. Generally, the self-diffusion coefficient is linked to the velocity-autocorrelation function (VACF) by a Green-Kubo relation. It is well established that the VACF displays an algebraic power-law decay \(t^{-3/2}\) in three-dimensional systems for underlying Newtonian dynamics . Similar anti-correlations with a tail \(t^{-5/2}\) are a general feature of strongly interacting systems without momentum conservation due to particle conservation. Only recently, these tails have been corroborated using Brownian dynamics simulations and MCT . We start with a short summary of the relevant MCT equations for the dynamics of a tracer particle, which resembles the collective equations . In the remaining part of we present the relevant equations for further tagged-particle quantities, in particular the MSD and the VACF. Additionally, the long-time behavior is discussed. The results within the framework of MCT are qualitatively compared with computer simulations in . We conclude in with a critical evaluation of our findings.
# Mode-coupling theory {#sec:theoretical_description}
## Tagged-particle motion
In this paper we investigate the dynamics of a tracer (tagged) particle with diameter \(\sigma_s\) and bare diffusion coefficient \(D_0^{(s)}\) suspended in a 3D quasi-confined colloidal liquid of identical hard spheres with diameter \(\sigma\) and bare diffusion coefficient \(D_0\). All particles evolve according to Brownian dynamics, where hydrodynamic interactions are ignored and are restricted to a quasi-confined environment by applying periodic boundary conditions along the confining direction, which we refer to as the \(z\)-direction. Thus, the transversal coordinates are folded back to \(-L/2\le z\le L/2\). The in-plane coordinates are denoted by \(\vec{r}=(x,y)\) and are infinitely extended in the thermodynamic limit, \(N\to\infty\), \(A\to\infty\), where \(N\) is the number of particles and \(A\) is the area in the dimensions perpendicular to the confining direction. We introduce the area density \(n_0=N/A\) and the volume density \(n=n_0/L\) related to the packing fraction \(\varphi=n\pi\sigma^3/6\) of the system. All coordinates and quantities of the tracer particle are labeled with a superscript \(s\). In this section, we convert the relevant MCT equations from the more general class of layered colloidal liquids to quasi-confined fluids. For this purpose we basically follow reference in which the corresponding collective equations have been derived. Since most of the steps of the derivation are identical if collective quantities are replaced by the corresponding single-particle quantities, we just provide a short summary of the main equations, focusing on the relevant differences to the collective case. The most relevant correlation function is the SISF, also called the incoherent intermediate scattering function, \[\begin{aligned}
\label{eq:SISF} S_{\mu}^{(s)}(q,t)=\langle \rho_\mu^{(s)}(\vec{q},t)^*\rho_\mu^{(s)}(\vec{q})\rangle, \end{aligned}\] given in terms of the fluctuating density modes \[\begin{aligned}
\rho_\mu^{(s)}(\vec{q})=\exp{\left[\rmi Q_\mu z_s\right]}e^{\rmi \vec{q}\cdot\vec{r}_s}, \end{aligned}\] of the tagged particle. The initial value \(S_\mu^{(s)}(q,t=0)=1\) is independent of \(q\) and \(\mu\). The two-dimensional wavevector \(\vec{q}=(q_x,q_y)\) in the lateral direction becomes continuous in the thermodynamic limit, whereas the wavenumber \(Q_{\mu}=2\pi\mu/L,\mu\in\mathbb{Z}\) is discrete. For \(L\to\infty\) rotational invariance with the 3D wavevector \(\vec{k}=(\vec{q},Q_\mu)\) is restored and the bulk limit is reached as it has been demonstrated for the general class of layered fluids . Due to periodic boundary conditions, the accessible \(z\)-coordinate positions do not differ between the tracer and the host-liquid particles, even if the tracer particle varies in size. This is in contrast to the case of a tracer particle confined between two parallel walls, where the effective confinement length \(L_s\) perpendicular to the walls for the tracer particle does not have to coincide with the accessible slit width \(L\) for the host-liquid particles . Therefore, for the tagged particle within the slit geometry an additional plane wave basis \(\exp{(-\rmi Q_\mu^{(s)}z)}\), with \(Q_{\mu}^{(s)}=2\pi\mu/L_s,\mu\in\mathbb{Z}\) is required. Quasi-confinement, however, simplifies the situation and no additional plane wave basis is necessary. The exact equations of motion (e.o.m.) for \(S_{\mu}^{(s)}(q,t)\) can be derived using the Zwanzig projection operator formalism introducing an effective self-memory kernel \(M_\mu^{(s)}(q,t)\) analogous to the collective case \[\begin{aligned}
\label{eq:eom_tagged} \fl\dot{S}_\mu^{(s)}(q,t)+&D_\mu^{(s)}(q)S_\mu^{(s)}(q,t) + D_\mu^{(s)}(q)\int_0^t M_\mu^{(s)}(q,t-t^\prime)\dot{S}_\mu^{(s)}(q,t^\prime)\mathrm{d} t^\prime = 0, \end{aligned}\] with initial decay \(D_\mu^{(s)}(q)=D_0^{(s)}\left(q^2+Q_\mu^{2}\right)\) related to the tagged-particle bare diffusion coefficient \(D_0^{(s)}\). Due to the quasi-confined geometry the relaxation naturally splits into channels parallel and perpendicular to the confining direction. Following the same procedure as for the collective case, the effective kernel is related to the matrix-valued irreducible memory kernel \(\boldsymbol{\mathcal{M}}_\mu^{(s)}(q,t)\) with diagonal components \(\mathcal{M}_\mu^{\parallel\parallel,(s)}(q,t)\) and \(\mathcal{M}_\mu^{\perp\perp,(s)}(q,t)\) and the symmetric off-diagonal element \(\mathcal{M}_\mu^{\parallel\perp,(s)}(q,t)\) by the relation \[\begin{aligned}
\label{eq:effective_memory_time_tagged} D_{\mu}^{(s)}(q) M_{\mu}^{(s)}(q,t) &+ D_0^{(s)2} \int_0^t M_{\mu}^{(s)}(q,t-t^\prime)\alpha_{\mu}^{(s)}(q,t^\prime)\mathrm{d}t^\prime = D_0^{(s)} \beta_{\mu}^{(s)}(q,t) \nonumber\\ &+ D_0^{(s)2}\int_0^t \mathcal{M}_{\mu}^{\parallel\parallel,(s)}(q,t-t^\prime) \mathcal{M}_{\mu}^{\perp\perp,(s)}(q,t^\prime)\mathrm{d}t^\prime \nonumber \\ &-D_0^{(s)2}\int_0^t \mathcal{M}_{\mu}^{\parallel\perp,(s)}(q,t-t^\prime)\mathcal{M}_{\mu}^{\parallel\perp,(s)}(q,t^\prime)\mathrm{d}t^\prime. \end{aligned}\] Therein, the coefficients \[\begin{aligned}
\fl\alpha_{\mu}^{(s)}(q,t) = Q_{\mu}^{2}\mathcal{M}_{\mu}^{\parallel\parallel,(s)}(q,t) + q^2\mathcal{M}_{\mu}^{\perp\perp,(s)}(q,t)-2 q Q_{\mu}\mathcal{M}_{\mu}^{\parallel\perp,(s)}(q,t), \end{aligned}\] and \[\begin{aligned}
\fl\beta_{\mu}^{(s)}(q,t) = \frac{q^2}{q^2+Q_{\mu}^{2}} \mathcal{M}_{\mu}^{\parallel\parallel,(s)}(q,t) + \frac{Q_{\mu}^{2}}{q^2 + Q_{\mu}^{2}}\mathcal{M}_{\mu}^{\perp\perp,(s)}(q,t) + \frac{2q Q_{\mu}}{q^2+Q_{\mu}^{2}}\mathcal{M}_{\mu}^{\parallel\perp,(s)}(q,t), \end{aligned}\] are completely determined by the matrix entries of the irreducible memory kernel \(\boldsymbol{\mathcal{M}}_\mu^{(s)}(q,t)\). Finally, MCT approximates the irreducible memory kernel as a bilinear functional of the ISF and the SISF, where the coupling between different modes and wavenumbers is described by vertices. These are entirely determined by the direct correlation function \(c_\mu^{(s)}(q)\) of the tagged particle, which is connected to the static structure factor \(S_\mu^{(s)}(q)\) by a generalized Ornstein-Zernike relation . Numerical efficiency can be remarkably enhanced by introducing bipolar coordinates \[\begin{aligned}
\fl\mathcal{M}_\mu^{\alpha\beta,(s)}(q,t) = \frac{n}{4L^3\pi^2}\int_0^\infty\int_{|q-q_1|}^{q+q_1} & \sum_{\substack{\mu_1=-\infty\cr \mu_2=\mu-\mu_1}}^{\infty} S_{\mu_1}(q_1,t) S_{\mu_2}^{(s)}(q_2,t)\frac{4q_1q_2} {\sqrt{4q^2q_1^2-(q^2+q_1^2-q_2^2)^2}} \nonumber \\ &\times\left[b^\alpha\left(\frac{q^2+q_1^2-q_2^2}{2q},Q_{\mu_1}\right)c_{\mu_1}^{(s)}(q_1)\right] \nonumber\\ & \times\left[b^\beta\left(\frac{q^2+q_1^2-q_2^2}{2q},Q_{\mu_1}\right)c_{\mu_1}^{(s)}(q_1)\right]\mathrm{d} q_1\mathrm{d} q_2, \end{aligned}\] where \(b^\alpha(x,z)=x\delta_{\alpha \parallel}+z\delta_{\alpha \perp}\) is the selector with respect to the channel indices \(\alpha, \beta \in \{\parallel,\perp\}\). The indices \(\mu_1\) and \(\mu_2\) run over all integer values satisfying the selection rule \(\mu=\mu_1+\mu_2\), which is valid due to translational invariance along the confining direction. Since not only the SISF, but also the ISF enter the equations and all tagged-particle quantities within MCT are derived quantities, the collective dynamics have to be solved first. The algorithm for the tagged-particle dynamics is equivalent to the collective case. Solving the coupled equations for the effective (self-)memory kernel and the (S)ISF is not straightforward. A detailed description of the algorithm using a decimation scheme can be found in references . To distinguish ergodic liquid states from arrested localized states it is useful to introduce the nonergodicity parameter or glass-form factor as the long-time limit of the SISF, \[\begin{aligned}
F_\mu^{(s)}(q) := \lim_{t\to\infty} S_\mu^{(s)}(q,t). \end{aligned}\] Liquid states then correspond to vanishing nonergodicity parameters, whereas \(F_\mu^{(s)}(q)\) remains finite for glassy states. The nonergodicity parameters are directly accessible in simulations or experiments, and within the framework of MCT \(F_\mu^{(s)}(q)\) can be determined without solving the full time-dependent equations explicitly .
## Long-wavelength limit
We now elaborate only the lateral dynamics of the quasi-confined liquid from a "top-view" perspective and investigate the mean-square displacement (MSD) and the velocity-autocorrelation function (VACF). Even though no confinement is present in the parallel direction, the dynamics of this two-dimensional projection are still sensitive to the confining length. To derive explicit equations for the MSD and the VACF we investigate the tagged-particle dynamics in the long-wavelength limit. The small \(q\) series of the SISF \[\begin{aligned}
\label{eq:SISF_small_q} S_\mu^{(s)}(q,t)=P_\mu(t)-q^2 \delta r_\mu^{2} (t)/4 + \mathcal{O}(q^4), \end{aligned}\] together with the out-of-plane self-intermediate scattering function \[\begin{aligned}
\label{eq:P_mu} P_\mu(t)=\left\langle\exp{\left[-\rmi Q_\mu\Big(z(t)-z(0)\Big)\right]}\right\rangle, \end{aligned}\] can be written in terms of a generalized MSD \[\begin{aligned}
\delta r_\mu^2(t) := \left\langle\exp\left[-\rmi Q_\mu\Big(z(t)-z(0)\Big)\right]\left[\vec{r}(t)-\vec{r}(0)\right]^2\right\rangle. \end{aligned}\] \(P_\mu(t)\) does not depend on the in-plane wavevector \(\vec{q}\), and probes the dynamics of the tracer only along the confinement direction. Inserting the expansion for \(S_0^{(s)}(q,t)\) with \(P_0(t)=1\) in the e.o.m. for the tagged-particle correlator, equation , we obtain for the MSD \[\begin{aligned}
\delta r^2_{0}(t)+D_0^{(s)}\beta^2\int_0^t \zeta^{(s)}_{0}(t-t^\prime)\delta r^2_{0}(t^\prime)\mathrm{d} t^\prime = 4 D_0^{(s)} t, \end{aligned}\] reminiscent of the corresponding e.o.m. in 2D systems . However, in the case of quasi-confinement, a second equation for the long-wavelength limit of the effective self-memory kernel \(\beta^2\zeta_0^{(s)}(t)=\lim_{q\to 0}q^2M_0^{(s)}(q,t)\) is necessary. The irreducible memory kernel \(\boldsymbol{\mathcal{M}}_\mu^{(s)}(q,t)\) for arbitrary mode index \(\mu\) decouples in the long-wavelength limit, \(\lim_{q\to 0}\mathcal{M}_\mu^{\alpha\beta,(s)}(q,t)=\mathcal{M}_\mu^{\alpha,(s)}(t)\delta_{\alpha\beta}\). The non-vanishing diagonal components are then given by \[\begin{aligned}
\label{eq:memory_parallel} \mathcal{M}_\mu^{\parallel,(s)} (t) = \frac{n}{4\pi L^3} \int & k^3 \mathrm{d} k\,\sum_{\mu_1}\left(c_{\mu_1}^{(s)}(k)\right)^2 S_{\mu_1}(k,t) S_{\mu-\mu_1}^{(s)}(k,t), \end{aligned}\] and \[\begin{aligned}
\label{eq:memory_perp} \mathcal{M}_\mu^{\perp,(s)} (t) = \frac{n}{2\pi L^3} \int & k\mathrm{d} k\, \sum_{\mu_1}\left(Q_{\mu_1}c_{\mu_1}^{(s)}(k)\right)^2 S_{\mu_1}(k,t) S_{\mu-\mu_1}^{(s)}(k,t), \end{aligned}\] enabling us to determine the equation for the effective memory kernel \(\beta^2\zeta_0^{(s)}(t)\) in the long-wavelength limit \[\begin{aligned}
\label{eq:effective_memory} D_0^{(s)} \beta^2\zeta_0^{(s)}(t) &+ D_0^{(s)2} \beta^2 \int_0^t \zeta_0^{(s)}(t-t^\prime)\mathcal{M}_0^{\perp,(s)}(t^\prime)\mathrm{d} t^\prime = D_0^{(s)} \mathcal{M}_0^{\parallel,(s)} (t) \nonumber\\ &+ D_0^{(s)2} \int_0^t \mathcal{M}_0^{\parallel,(s)}(t-t^\prime)\mathcal{M}_0^{\perp,(s)}(t^\prime)\mathrm{d} t^\prime. \end{aligned}\] The equation for the effective memory kernel is not restricted to \(\mu=0\) relevant for the MSD, but can also be extended to arbitrary mode indices \[\begin{aligned}
D_0^{(s)} Q_\mu^{2} M_\mu^{(s)}(t) &+ D_0^{(s)2}\int_0^t Q_\mu^{2} M_\mu^{(s)}(t-t^\prime)\mathcal{M}_\mu^{\parallel,(s)}(t^\prime)\mathrm{d} t^\prime = D_0^{(s)} \mathcal{M}_\mu^{\perp,(s)}(t) \nonumber\\ &+ D_0^{(s)2}\int_0^t \mathcal{M}_\mu^{\parallel,(s)}(t-t^\prime)\mathcal{M}_\mu^{\perp,(s)}(t^\prime)\mathrm{d} t^\prime, \quad\mu\neq 0, \end{aligned}\] with the long-wavelength limit \(M_\mu^{(s)}(t)=\lim_{q\to 0}M_\mu^{(s)}(q,t)\) for higher order effective self-memory kernels. From the expansion for \(S_\mu^{(s)}(q,t)\), equation , we get \[\begin{aligned}
\dot{P}_\mu(t)&+D_0^{(s)}Q_\mu^2 P_\mu(t) +D_0^{(s)}\int_0^t Q_\mu^2 M_\mu(t-t^\prime)\dot{P}_\mu(t^\prime)\mathrm{d} t^\prime=0, \end{aligned}\] as the e.o.m. for \(P_\mu(t)\). In contrast to the MSD, which only considers the dynamics in the parallel direction, it takes into account the dynamics perpendicular to the confining direction. The mathematical structure of this integro-differential equation is reminiscent of the e.o.m. for the tagged-particle correlator, equation , and therefore it is straightforward to solve it numerically. The velocity-autocorrelation function (VACF) \(Z_0(t)\) of a quasi-confined colloidal liquid can formally be connected to the MSD by the relation \[\begin{aligned}
\label{eq:VACF} Z_0(t) := \frac{1}{4}\frac{\mathrm{d}^2}{\mathrm{d} t^2}\delta r^2_0(t). \end{aligned}\] Therefore, the in-plane VACF can in principle be determined by taking derivatives of the MSD. Yet, due to discretization of the wavenumber grids in MCT, long-wavelength phenomena cannot be properly resolved. An exact e.o.m. can be readily derived using the Zwanzig projection operator formalism with the same irreducible effective memory kernel \(\beta^2\zeta_0^{(s)}(t)\) as already introduced above in the e.o.m. for the MSD \[\begin{aligned}
\label{eq:eom_VACF} Z_0(t)+ D_0^{(s)}\beta^2\int_0^t \zeta_0^{(s)}(t-t^\prime)Z_0(t^\prime)\mathrm{d} t^\prime =-D_0^{(s)2}\beta^2\zeta_0^{(s)}(t). \end{aligned}\]
## Long-time behavior
We elaborate the long-time behavior of the VACF within the framework of MCT explicitly. For long times, the integrals for the irreducible memory kernel, equations and , are dominated by small wavenumbers. In the long-wavelength limit, \(\mu=0\) is the only mode which becomes slow and contributes to the sum. Thus, the perpendicular component vanishes, \(\mathcal{M}_0^{\perp,(s)}(t)\simeq 0\). Correspondingly, for long times the equation for the effective memory kernel, equation , reduces to \(D_0^{(s)}\beta^2\zeta_0^{(s)}(t) = D_0^{(s)} \mathcal{M}_0^{\parallel,(s)}(t)\). Approximating the ISF and SISF for \(\mu=0\) by collective diffusive motion, \(S_0(q,t)\approx S_0(0)\exp\left(-D_0 q^2 t /S_0(0)\right)\) and self-diffusion \(S_0^{(s)}(q,t)\approx \exp\left(-D_{0}^{L,(s)}q^2t\right)\), the long-time behavior of the VACF follows from the equation of motion, equation , using Tauber's theorem \(Z(t)\simeq-(D_0^{L,(s)})^2 \beta^2\zeta_0^{(s)}(t)\). The explicit representation of the long-time asymptote reads \[\begin{aligned}
\label{eq:VACF_LTT} Z_0(t) \simeq-\left(D_0^{L,(s)}\right)^2 \frac{n}{4\pi L^3} \frac{\left(c_{0}^{(s)}(0)\right)^2S_{0}(0)}{2\left(D_0^{L,(s)}+D_0/S_0(0)\right)^2} t^{-2}. \end{aligned}\] This power-law tail \(\sim-t^{-2}\) is similar to the persistent anti-correlations found in bulk systems . Due to the confinement, long-wavelength phenomena are only important along the lateral direction and we identify for long times the expected power-law for a two-dimensional system with a different prefactor. In principle, autocorrelation functions similar to the VACF can also be defined for higher order modes, \(\mu \ge 1\), characterizing the out-of-plane dynamics with the same approach as for the MSD. However, in this work we focus on the long-time behavior of the VACF, since no long-time tails are expected for \(\mu\ne 0\).
# Simulations and numerical results {#sec:results}
After describing the set-up for the computer simulations, we analyze the tagged-particle quantities (self-nonergodicity parameters, SISF, MSD and VACF) from a polydisperse hard-sphere fluid and compare them with our results from MCT. The long-time behavior of the VACF is solely analyzed within the framework of MCT. The tracer particle is identical to the bath particles with the same diameter \(\sigma\) and bare diffusion coefficient \(D_0\). Accordingly, it becomes particularly simple to improve statistics in the simulation results. As input for the MCT calculations we use the static structure factors \(S_\mu(q)\) and direct correlation functions \(c_\mu(q)\) elaborated within liquid state theory using the Percus-Yevick (PY) closure relation. A detailed comparison between PY and monodisperse simulations is done in reference and a short assessment of the differences between PY and polydisperse simulation data has been performed recently in connection with the discussion of the collective dynamics in quasi-confined systems .
## Simulations
We perform event-driven molecular dynamics simulations for a 3D hard-sphere system, with periodic boundary conditions in all directions. The quasi-confinement is realized by setting the length of the \(z\)-direction of the simulation box much smaller than the other two directions. The extension of the unconfined directions is \(25\sigma\), whereas the confining length, \(L\), has only a few particle diameters. Although the simulated particles are driven by Newtonian dynamics, we expect that the qualitative behavior, in particular for long times, will be comparable to MCT results for colloidal liquids undergoing Brownian motion . To prevent crystallization polydisperse particles are used in the simulations. The particle diameters are given by an inverse-occupied volume distribution with a standard deviation of \(0.117\sigma\), setting the polydispersity to \(11.7\%\). It must be ensured that particles are initialized without overlap, which we do here by initializing particles in a larger volume, and compressing the system with standard algorithms to reach the desired packing fraction and system size . The time scale of the simulation is determined by the thermal energy \(k_BT\) and the particle mass \(m\), where we define \(t_0=\sqrt{m\sigma^2/k_BT}\). The SISF can directly be calculated from the density-mode representation, equation . In general, statistics for incoherent quantities can be significantly enhanced compared to collective ones by averaging over all particles. As we have discussed previously , it is not practical to reach a fully equilibrated glassy fluid of polydisperse hard spheres in quasi-confinement, because as in the slit geometry , the particles will eventually demix and then crystallize. However, this process occurs on long timescales, outside what is experimentally relevant, and only accessible in simulations when relying on advanced, unphysical dynamical algorithms. Therefore, to study glassy dynamics here, we use Newtonian dynamics only, restrict ourselves to relatively short simulation times, and carefully check the results for any effects of aging. Our data is taken over a time \(10^5 t_0\), after a sample preparation time \(10^5 t_0\). We have also repeated all simulations using a longer sample preparation time of \(10^6 t_0\), and found this does not change the qualitative behavior we will discuss in this section, and therefore the conclusions drawn are not affected by aging. The simulation data for the SISF can be approximately described by the phenomenological Kohlrausch-William-Watts (KWW) stretched exponential \[\begin{aligned}
\mbox{KWW fit: } S_\mu^{(s)}(q,t)=F_\mu^{(s)}(q)\exp\left\{-\left[t/\tau_\mu^{(s)}(q)\right]^{\beta_\mu^{(s)}(q)}\right\}. \end{aligned}\] The data is fit in the time range \(t\in(10 t_0,10^5 t_0)\) and the incoherent quantities, namely the Kohlrausch exponent \(\beta_\mu^{(s)}(q)\), the relaxation time \(\tau_\mu^{(s)}(q)\) and the self-nonergodicity parameter \(F_\mu^{(s)}(q)\), can be extracted.
## Self-nonergodicity parameter
Unless otherwise stated, the numerical results are obtained within MCT working on an equidistant wavenumber grid \(q=\hat{q}\Delta q + q_0\) with grid points \(\hat{q}=0,\dots,N_q-1\) parallel to the confinement. As input parameters \(q_0\sigma=0.1212\), \(\Delta q \sigma = 0.404\) and \(N_q=100\) are used. The discrete mode index associated with the confining direction is limited to \(|\mu|\le 15\). We assume that the tagged particle is of the same species as the bath particles, therefore the identical direct correlation functions as in the collective case enter the MCT equations. We start by analyzing the non-monotonic behavior of the nonergodicity parameters depending on the confinement length. We choose three different lengths \(L=2.0\sigma, 2.5\sigma, 2.8\sigma\) where the effects are most pronounced. The non-monotonic behavior coincides qualitatively with the collective case , therefore we believe that discussing more different values does not contribute to a deeper understanding. Since it is known that MCT underestimates the critical packing fraction not only in the bulk, but also for confined systems , MCT results for \(\varphi=0.53\) ( (dashed lines)) are compared to simulations with a higher packing fraction \(\varphi=0.59\) ( (solid lines)), therefore the dynamics are glassy in both cases. In principle, this discrepancy in the packing fraction can be reduced with modified MCT approaches , but it does not affect the validity of our conclusions in this work. Referring to the nonequilibrium-state diagram of quasi-confined liquids , MCT results are presented for \(\varphi=0.53\) in contrast to the remainder of this manuscript to ensure that the behavior is glassy for all investigated confinement lengths. Within MCT, the bell-shaped curves manifest non-monotonic behavior for all wavenumbers. The incommensurate value \(L=2.5\sigma\) shows a higher structural arrest compared to the more commensurate values \(L=2.0\sigma\) and \(L=2.8\sigma\). The confinement length is referred as commensurate if the ratio \(L/\sigma\) is close to an integer and as incommensurate if it is close to a half-integer. The non-monotonic behavior reflects the collective behavior in quasi-confinement and is also quite similar to the incoherent nonergodicity parameter within the slit geometry , although the non-monotonic behavior is less noticeable due to missing layering. In contrast, for the simulations non-monotonic behavior is not obvious. Indeed, the structural arrest for \(L=2.5\sigma\) is higher compared to \(L=2.0\sigma\), but it is almost indistinguishable from \(L=2.8\sigma\), where we expect a significant drop in the structural relaxation. Using the von Schweidler law instead of KWW fits, a marginal difference can be observed. Yet, already in the slit case the non-monotonic behavior of the nonergodicity parameters is less pronounced for the simulations compared to the MCT results . Since effects are considerably smaller in quasi-confinement anyway, the overlap of the curves for \(L=2.8\sigma\) and \(L=2.8\sigma\) is not surprising. This behavior indicates that the height of the plateau value of the SISF does not show a pronounced non-monotonic behavior, which is most likely due to the inclusion of polydispersity, in contrast to the MCT predictions and simulation data for the collective motion . It has been shown for bulk systems that the incoherent nonergodicity parameter is less sensitive to effects of different particle sizes of the constituents compared to the collective case. Therefore, we conclude that also effects of local order between commensurate and incommensurate packing have a minor influence on the incoherent nonergodicity parameter. For both MCT and simulations the lowest order mode \(F_0^{(s)}(q)\) reaches unity in the long-wavelength limit reflecting particle conservation in the system. For higher order modes, \(F_\mu^{(s)}(q\to 0)\) corresponds to \(P_\mu(t\to\infty)\) according to equation , with a value different from unity in agreement with equation . \(P_\mu(t)\) is only sensitive along the confining direction, and for large confinement length the nonergodicity parameter should approach the bulk limit. Then, it only depends on the magnitude of the 3D wave vector \(k=\sqrt{q^2+Q_\mu^{2}}\) and rotational invariance is restored, \(F_\mu^{(s)}(q)\to F^{(s)}(k)\). The drop between \(F_0^{(s)}(q\to 0)\) and \(F_1^{(s)}(q\to 0)\) can be observed in for simulations, whereas for MCT it becomes only visible for higher order modes, \(\mu\ge 2\), not included in the figure. Apart from the long-wavelength behavior the appearance of \(F_1^{(s)}(q)\) is very similar to \(F_0^{(s)}(q)\) indicating that the zero-order static quantities \(c_0(q)\) and \(S_0(q)\) are also the relevant quantities for the higher order self-nonergodicity parameters, since the static quantities \(c_\mu(q)\) and \(S_\mu(q)\) are significantly different between \(\mu=0\) and \(\mu=1\). Due to the periodic boundary conditions and the absence of layering, differences between the separate modes are less significant compared to the slit case .
## SISF
We proceed with the full dynamic solution of the SISF. The unit of time for the MCT results is set by \(\sigma^2/D_0\). (a) shows the first two modes \(S_0^{(s)}(q,t)\) and \(S_1^{(s)}(q,t)\) (inset) for MCT at packing fraction \(\varphi=0.515\), where it is known from the nonequilibrium-state diagram that a reentrant behavior occurs . The related data for simulations at the higher packing fraction \(\varphi=0.59\) are displayed in (b). The MCT results look quite similar to the corresponding collective ISF with clearly observable non-monotonic behavior depending on the confinement length. The MCT solution either decays to zero for the more commensurate values (\(L=2.0\sigma\), \(L=2.8\sigma\)), identifying liquid states, or it reaches a non-vanishing plateau value corresponding to an arrested glassy state for the incommensurate value (\(L=2.5\sigma\)). The increased longitudinal diffusion in the case of commensurate packing prevents slowing down of the dynamics, and therefore, the arrested glassy state is shifted to higher packing fractions compared to incommensurate packing . The behavior of the first higher-order mode \(S^{(s)}_1(q,t)\) almost duplicates the one of \(S^{(s)}_0(q,t)\). Whereas \(S^{(s)}_0(q,t)\) only considers the dynamics parallel to the confining direction, higher order modes of the SISF also include the perpendicular direction. The similarity between the modes of the SISF suggests that not only the long-time limits, but the whole dynamics are mainly driven by the zero-order mode of the static quantities, \(S_0(q)\) and \(c_0(q)\), in particular by the first sharp diffraction peak, like in the collective case . Notably, \(S_1^{(s)}(q,t)\) decays faster than \(S_0^{(s)}(q,t)\) for all studied lengths. In close resemblance to the collective ISF, the simulation data only show an intermediate plateau for all studied lengths, accompanied by a stretched relaxation. The general shape of the SISF coincides with the collective ISF with non-monotonic dependence on the confinement length. Yet, there are minor differences compared to the ISF, in particular the precise plateau values (incoherent nonergodicity parameters) differ as it has been discussed in the preceding section. Even though the non-monotonic dependence is not obvious in the plateau value, it still can be observed in the corresponding wavenumber dependent relaxation times \(\tau_0^{(s)}(q)\) ( (a)) and the Kohlrausch exponents \(\beta_0^{(s)}(q)\) ( (b)). It would be interesting to analyze if the large-wavenumber limit of the Kohlrausch exponent coincides with the von Schweidler exponent similar to other recent studies comparing the von Schweidler and the KWW equation . Unfortunately, we cannot accurately determine the von Schweidler exponent form our data. Even for higher order modes, where the SISF almost overlap (inset of (b)), the related relaxation times \(\tau_1^{(s)}(q)\) and Kohlrausch exponents \(\beta_1^{(s)}(q)\) reveal considerable non-monotonic behavior ( (insets)). Therefore, we conclude that within the simulations the non-monotonic dependence is absent to a large extend in the plateau heights, but it is still present in the dynamics manifested by the relaxation times and the Kohlrausch exponents. Simulation data also confirms that higher modes of the SISF decay faster than the zero order quantity.
## MSD {#sec:results_msd}
The mean-square displacement (MSD) is calculated for different confining lengths \(L\) at the same packing fractions \(\varphi=0.515\) for MCT ( (a)) and \(\varphi=0.59\) for simulations ( (a)) as above for the SISF. Quantitative differences in the limit of small \(t\) arise due to different microscopic dynamics. Whereas the MCT equations are derived for Brownian microscopic dynamics characterizing colloidal liquids, the simulations rely on Newtonian molecular dynamics. Therefore, for MCT, diffusive motion \(\delta r^2_0(t)\simeq 4D_0 t\), \(t\to 0\) can be observed for all investigated confinement lengths in the limit of small \(t\), in contrast to the simulations, which show ballistic motion. With increasing time the dynamics are slowed down in the crossover regime as has already been discussed in connection with the SISF. For even larger times, particles are trapped due to the cage effect in the glassy state (\(L=2.5\sigma\)), which becomes apparent in the MCT solution by a plateau value of the MSD. In contrast, the data for the other confinement lengths show liquid-like behavior with long-time diffusion \(\delta r^2_0(t)=4D_0^{L,(s)} t\), \(t\to\infty\) after the crossover regime with a two-step relaxation process distinctive for MCT dynamics. This second rise is the \(\alpha\)-process of the MSD, indicating the tagged-particle's exit of the cage. For simulations, particles move out of the cage eventually for all confinement lengths, and diffusive motion can be observed at long times. This behavior is consistent with the intermediate nature of the plateau in the SISF. The inset of (a) shows the non-monotonic behavior of \(P_1(t)\) within MCT. The general shape of the relaxation is quite similar compared to the SISF for small \(q\). The curves are indistinguishable for small times \(tD_0/\sigma^2 \lesssim 10^{-1}\) before the crossover regime sets in. All liquid curves decay to zero and the finite value for glassy states corresponds to \(F^{(s)}_1(q\to 0)\) as already mentioned above. A characteristic localization length \(r_0^{(s)}\) can be defined by the small-\(q\) behavior of the nonergodicity parameter, \(F_0^{(s)}(q)=1-\left(q r_0^{(s)}\right)^2 + \mathcal{O}(q^4)\) similar to the bulk case , which is related to the long-time limit of \(\delta r_0^2(t)\) by \[\begin{aligned}
\lim_{t\to\infty} \delta {r}_0^{2} (t) = 4 r_0^{(s)2}. \end{aligned}\] In the glassy state (\(L=2.5\sigma\) in ) the finite value of the localization length \(r_0^{(s)2}=\lim_{t\to\infty} \delta {r}_0^{2} (t)/4\) related to the MSD corresponds to the size of cages in which the colloidal particles are trapped and reflects the behavior of colloidal suspensions in experiments for the slit geometry , yet no physical boundaries are present in our quasi-confined system. indicates that the product of the long-time self-diffusion coefficient \(D_0^{L,(s)}\) and the self-relaxation time \(\tau_0^{(s)}\) at the structure factor peak, \(q\sigma=7.0\), is approximately constant, albeit the individual contributions vary significantly (by a factor of 50 for MCT and 3 for simulations respectively) with the confinement length. Within MCT, we identify the relaxation time \(\tau_0^{(s)}\) with the time where the SISF has reduced to \(1/e\). Since the dynamics are glassy within MCT for \(L=2.5\sigma\) no self-relaxation time and self-diffusion coefficient can be extracted. MCT predicts a single \(\alpha\)-relaxation scale, which implies the validity of the Stokes-Einstein (SE) relation for straight paths approaching the glass-transition line. However, since a reentrant phenomenon occurs between \(L=2.0\sigma\) and \(L=2.8\sigma\), the SE relation does not have to hold here. Nevertheless, it is approximately fulfilled, albeit a slightly non-montonic behavior of \(\tau_0^{(s)}D_0^{L,(s)}\) within the simulations. It is known that the violation of the SE relation sets in close to the critical packing fraction.
:::
## VACF {#sec:results_vacf}
Due to the dominating linear growth of diffusion some subtleties are usually hidden in the MSD, therefore, it is instructive to also investigate the VACF. shows the VACF \(Z_0(t)\) extracted from simulations for different confinement lengths. The prominent minimum in the VACF is related to the strong caging of particles for intermediate time scales. Quite similar curve shapes are present in the glass-like behavior of confined supercritical Argon with inhomogeneous density profiles. As can be seen, non-monotonic behavior for short and intermediate times is only marginally pronounced, but still present (lower left inset). Since the crossover regime within the MSD for liquid states () extends over several orders of magnitude in time and just appears as an interpolation between the short-time diffusive motion \(\delta r^2_0(t)=4D_0^{(s)} t, t\to 0\) and the long-time diffusion \(\delta r^2_0(t)=4D_0^{L,(s)} t,t\to \infty\), we particularly want to investigate correlations of the VACF in the long-time regime. Unfortunately, due to insufficient statistics for the simulation data it is not possible to study long-time effects for the simulations, and we limit our discussion to MCT results. In principle, advanced simulation algorithms for colloidal bulk liquids have been developed recently to investigate the long-time behavior also for simulations . The numerical evaluation within MCT is not straightforward and we have to rely on a logarithmic wavenumber grid similar to the 3D bulk case to extract the long-time behavior. The reduced wavenumbers are then given by \([x^{-N},\dots,x,1]q_{\mbox{\scriptsize{max}}}\sigma\) with high-\(q\) cutoff \(q_{\mbox{\scriptsize{max}}}\sigma = 40\), base \(x=1.02\) and \(N=400\) grid points. This corresponds to a minimal wavenumber \(q_{\mbox{\scriptsize{min}}}\sigma=0.0148\), ensuring that the long-time behavior of the VACF is properly resolved. Additionally, only \(|\mu|\le 10\) modes are used to save computational resources. In principle, there are more sophisticated grid types which have been discussed lately for two-and three-dimensional systems . The main idea behind these nonuniform grids is to have a better representation with more grid points near the maxima and minima of the structure factor and the direct correlation function, but less wavenumbers in regions where the behavior is more linear. The implementation of this scheme for quasi-confined liquids would require a different wavenumber grid for each mode \(\mu\), since the extreme values of the static input sensitively depend on the wavenumber for different mode indices . Still, we find for the long-time behavior of the VACF it is sufficient to have a good long-wavelength resolution since the corresponding integrals are dominated by low wavenumbers instead of the structure factor peaks, and therefore, we simply rely on a logarithmic grid independent of the mode index. At the arbitrarily chosen confinement length \(L=2.0\sigma\) the numerical results for the reduced VACF \(-Z_0(t)/\varphi\) for MCT display non-trivial correlations beyond the crossover regime of the MSD with anti-correlated long-time tails \(Z_0(t)\sim-t^{-2}\) for all investigated packing fractions, see . Yet, in contrast to the bulk behavior, \(Z(t)\sim-t^{-5/2}\), the exponent indicates two-dimensional dynamics with the prefactor sensitively depending on the density, equation . Surprisingly, the non-monotonic dependence on the confinement length only becomes apparent for larger packing fractions \(\varphi\gtrsim 0.45\) approaching the glass-transition (inset). For this comparatively high packing fraction non-monotonic behavior depending on the confinement length can be observed in two respects. Firstly, the crossover regime is shifted to larger times for \(L=2.5\sigma\) compared to the other confining lengths and secondly, the tails are more suppressed for incommensurate values. This demonstrates that evidence of glassy-like behavior can be observed even at \(\varphi=0.47\), since these are the same values of \(L\) which correspond to glassy states at higher packing fraction. Indeed, the strong \(L^3\) dependence in the denominator of equation is compensated for by the \(L\)-dependent structural quantities \(S_0(0)\), \(c_0(q)\) and the long-time self-diffusion coefficient \(D_0^{L,(s)}\) if the density is not too high. For large confinement lengths (\(L\gtrsim 5.0\sigma\)) the dynamics of the particle are indistinguishable from bulk motion for short times with an intermediate window, where the VACF decays with a power-law, \(Z(t)\sim-t^{-5/2}\). Only for longer times, a crossover from the unconfined dynamics occurs and the expected power law for quasi-confinement, \(Z(t)\sim-t^{-2}\), is observed, similar to the tagged-particle within a slit . Extreme confinement, however, cannot be considered as the two dimensional limit case since due to periodic boundary conditions the hard spheres are still able to move perpendicular to the confining direction. Although, the long-time behavior in the simulations is covered by noise, indications of a similar intermediate tail \(Z(t)\sim-t^{-2}\) can be observed in the upper right inset of . Indeed, for much longer times the classical hydrodynamic Alder tails are expected for a momentum-conserving system of Newtonian hard-spheres. A comparable intriguing crossover from positive correlations to anti-correlations has been reported for dense hard-sphere systems . All MCT data show a divergent short-time behavior, \(Z(t)\sim-t^{-1/2}\) as \(t\to 0\), related to the hard-sphere potential. This divergence is not restricted to quasi-confined systems and can be observed in colloidal liquids with arbitrary dimension, e.g. in 3D bulk liquids . The exponent does not depend on the confining geometry, contrary to the long-time tail. MCT also gives a qualitative explanation for the strong variation of the long-time tail in terms of the static compressibility and slowing down of transport related to the 3D bulk case . In accordance with equation and reference the tail is suppressed by orders of magnitude with increasing \(\varphi\). In the dilute limit the self-diffusion coefficient in the denominator can be ignored and the prefactor of the long-time tail shows a direct dependence on the long-wavelength limit of the static structure factor, \(S_0(0)\). Approaching the glass transition the tails get even more suppressed due to the singularity of \(D_0^{L,(s)}\). At intermediate times we then expect to observe two additional power laws, which are directly related to the critical beta decay .
# Summary and conclusions {#sec:conclusion}
In this work, we have investigated the tagged-particle dynamics of quasi-confined colloidal liquids by mode-coupling theory and event-driven simulations. In particular, we have elaborated self-nonergodicity parameters, the self-intermediate scattering function (SISF), the mean-square displacement (MSD) and the velocity-autocorrelation function (VACF). By adapting our stable numerical algorithm to the self-dynamics we have found qualitative agreement between MCT and simulations, showing distinctive non-monotonic behavior with the confinement length. Within the simulations, the non-monotonic behavior in the SISF and incoherent nonergodicity parameters is slightly less pronounced compared to the collective case. Nevertheless, it is still present, particularly in the corresponding relaxation times. Since \(S_0(t)\) decays significantly more slowly than \(S_1(t)\) we conclude that the slowing down of the dynamics is most pronounced perpendicular to the confining direction. The behavior of the MSD nicely captures these features, and in the MCT clearly illustrates the transition from particle localization in the glassy state to particle diffusion in the liquid state. The non-monotonic behavior can be explained by an interplay between the local packing of the hard spheres and the confinement itself with an alternation between commensurate (sliding motion) and incommensurate packing (obstruction). Recent investigations between the local structure ordering and reentrance phenomena in a slit geometry suggest that local ordering changes from hexatic order in the commensurate case to local square order for incommensurate geometries . A similar approach in computer simulations tries to disentangle confinement and layering effects by applying an external potential to the unconfined system resulting in a similar density modulation as in the slit . Although the density profiles are quite similar in these two systems, they remain inhomogeneous in contrast to our quasi-confined liquid. The glass-transition line extracted from the nonequilibrium-state diagram separates liquid-like from non-ergodic states . Due to the smooth and continuous transition, we expect the generic \(A_2\) fold bifurcation in Arnol'd's terminology , although in principle higher-order singularities, where the critical scaling behavior at the \(A_3\) critical endpoint has been analyzed , are also possible within quasi-confined systems. In particular, these singularities should be present for competing mechanisms similar to bulk systems such as binary mixtures , attractive colloids , porous media or randomly pinned systems . To sum up, we have completed our recent study on quasi-confined liquids with investigating the tagged-particle dynamics. Even in the absence of layering in the density profile, there are pronounced non-monotonic effects, e.g. within the MSD, caused by the confinement itself. Generally, our system behaves similarly to strongly confined colloids within a slit . The MCT equations of the quasi-confined liquid are reminiscent of the diagonal approximation for the slit geometry. Therefore, a detailed MCT analysis can easily be adopted to quasi-confined liquids. Yet, we want to emphasize that for quasi-confined liquids it is an exact representation rather than an approximation. Besides, results between quasi-confinement and slit geometry always differ significantly due to important differences in the static input for the MCT equations. Generally, MCT and simulations coincide slightly better here in comparison to the slit geometry, e.g. manifested in the nonergodicity parameters (cf. reference ). This appears to be due to the homogeneous density profile in quasi-confined systems. In contrast, the pronounced layering within the slit is distorted in simulations by polydispersity. For the VACF we have mainly focussed on the long-time behavior identifying an anti-correlated algebraic decay for long times with an exponent characterizing the effective two-dimensional dynamics perpendicular to the confinement. These correlations are hidden in the MSD and can be rationalized within MCT. Future investigations could reveal similar tails within the non-Gaussian parameter or the Burnett coefficient . Our results are not limited to hard-sphere systems since several studies show the dynamic equivalence between soft colloids and hard sphere systems, e.g. recent experiments for soft star polymers . Non-monotonic confinement effects should also be present in driven systems like granular matter and active microrheology . This work has been supported by the Austrian Science Fund (FWF): I 2887. CFP gratefully acknowledges a Lise-Meitner fellowship of the Austrian Science Fund (FWF): M 2471. The computational results presented have been achieved in part using the HPC infrastructure LEO of the University of Innsbruck. | {'timestamp': '2021-09-14T02:22:01', 'yymm': '2109', 'arxiv_id': '2109.05497', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05497'} |
null | null |
null | null |
null | null |
null | null |
# Introduction
An *animal* is a planar shape with connected interior formed by gluing together a finite number of congruent regular polygons along their edges. Here, we study animals that have extremal combinatorial properties such as having maximally many shared (or interior) edges for a given number of polygons. We refer to animals attaining these optimal values as *extremal animals*. In the Euclidean plane, extremal animals have been well studied and it is known that they exhibit interesting geometric and combinatorial properties such as being isoperimetric, giving contact numbers for optimal disk packing configurations, and providing the right shape to produce animals with maximally many holes. In the hyperbolic plane, combinatorial and geometric properties of extremal animals have not been examined as closely. However, the relevance of extremal animals in this setting to other extremal combinatorial problems has already been established. These include optimal disk packing problems, the calculation of Cheeger constants, the establishment of the exponential growth constant of regular hyperbolic tessellations, and the implementation of algorithms to sample certain hyperbolic animals to approximate hyperbolic percolation thresholds. Recently, extremal hyperbolic animals have been used to model and study structures in applied mathematics, such as in crystallography theory for materials with hyperbolic symmetries, and in the architecture of circuit quantum electrodynamics. More generally, models based on hyperbolic tessellations have also been used in areas such as Bloch band theory in physics, and in quantum error correcting and quantum storage codes, among others. Therefore, having a better understanding of extremal hyperbolic animals might lead to interesting insights, results, techniques, and new research questions in these areas. The families of animals that we study in this paper are formed by finite subsets of tiles of regular Euclidean and hyperbolic tessellations. These tessellations can be parameterized using the Schläfli symbol \(\{p,q\}\), where \(p\) denotes the number of sides of the regular polygon forming the tessellation and \(q\) is the number of edges or tiles meeting at each vertex. It is well known that if \((p-2)(q-2)> 4\), \(=4\), or \(<4\), then the tessellation corresponds to the geometry of the hyperbolic plane, the Euclidean plane, or the sphere, respectively. We call an animal living in a \(\{p,q\}\)-tessellation a \(\{p,q\}\)-animal. The most well-known families of animals are those living in the Euclidean tessellations, namely for \(\{p,q\}\) either \(\{3,6\}\), \(\{4,4\}\), or \(\{6,3\}\). Formed by triangles, squares, or hexagons, respectively, these animals are also more commonly known as polyiamonds, polyominoes, and polyhexes in the literature. In 1976, Harary and Harborth studied extremal animals in the three Euclidean cases. More specifically, they gave algebraic expressions and bounds for the number of interior, exterior, and total numbers of both vertices and edges that an animal with a fixed number of tiles can have. Then they constructed a family of *spiral* animals which attain optimal values. Here, we extend Harary and Harborth's analysis to extremal hyperbolic animals. First, we define and analyze analogous sequences of *spiral* \(\{p,q\}\)-animals for \((p-2)(q-2) \geq 4\) and provide algebraic expressions for the number of vertices and edges in terms of a particular recursive sequence of integers. Then we prove that these *spiral* \(\{p,q\}\)-animals are indeed extremal. As a corollary of our results, we recover the formulas derived in for the Euclidean cases. We also explore enumeration of extremal hyperbolic animals. Exact enumeration of \(\{p,q\}\)-animals is only known for a few \(\{p,q\}\) pairs, and even then only for animals with a very small number of tiles. For instance, the number of fixed \(\{4,4\}\)-animals with \(n\) tiles is only known for \(n \leq 56\); and otherwise very little is known (see, for example, the entries of the OEIS A001207, A001420, A000228, A000577, A119611 for values in the \(\{3,6\}\)-, \(\{6,3\}\)-, and \(\{4,5\}\)-tessellations). In the particular case of Euclidean animals, there are also asymptotic results and computational approximations that give a better understanding of the problem of animal enumeration. In their closing thoughts in, Harary and Harborth raised the question of enumeration of extremal animals with \(n\) tiles. Surprisingly, this problem has remained largely unexplored, and it has been completely solved only for \(\{4,4\}\)-animals. Before now, nothing was known about the enumeration of extremal hyperbolic animals. As a consequence of our main results, for all \(\{p,q\}\) we find special sequences of extremal animals that are *unique* extremal animals, in the sense that any animal with the same number of tiles which is distinct up to isometries can't be extremal. .5cm
## Main Results
We begin by analyzing a sequence of layered animals, which we use as benchmarks for computing extremal values. These layered animals will attain extremal values, but more significantly they provide context for keeping track of how each subsequent tile affects all of the graph parameters in play. .75cm
For \(t>2\), we get as a corollary that the parameters of \(A_{p,q}(k)\) grow exponentially; as expected from the properties of hyperbolic geometry. The exponential growth rate of these parameters is governed by the constant \(\alpha>1\).
Some of the recursive formulas for the complete layered sequence \(A_{p,q}(k)\) in Theorem [\[thm:layers\]](#thm:layers){reference-type="ref" reference="thm:layers"} were independently computed and used as a tool for understanding geometric and combinatorial properties of the \(\{p,q\}\)-tessellations. However, the extremality of these values was not previously examined. For instance, in, Higuchi and Shirai use the layered \(A_{p,q}(k)\) structures to compute the Cheeger constant of the infinite graph of the \(\{p,q\}\)-tessellations. Specifically, they give recursive computations for \(\check{v}(k)\) and a value equivalent to the number of tiles in a layer, i.e., \(\check{n}(k)-\check{n}(k-1)\). Using these values they proved that the Chegeer constant of the \(\{p,q\}\)-tessellation, defined by \[\inf \left\{ \frac{|E(\partial_v W)|}{\text{vol}(W)}: W \text{ is a finite subgraph} \right\},\] is equal to the limit of the corresponding values for \(A_{p,q}(k)\) as \(k\to \infty\). Here, \(|E(\partial_v W)|\) is the number of edges connecting \(W\) with its complement, and vol\((W)=\sum_{x\in W} deg(x)\). The \(A_{p,q}(k)\) serve as a discrete analogue of balls of radius \(k\) centered at \(A_{p,q}(1)\), where we consider two faces to be adjacent if their intersection is non-empty. More commonly, proper combinatorial balls with respect to edge-distance in a graph with vertex set \(V\) are defined as \(\mathcal{B}_k(x)=\{y\in V: d(x,y)\leq k\}\). Then viewing the \(\mathcal{B}_k(x)\) as a structure on the dual graph of an animal, the definitions of \(A_{p,q}(k)\) and \(\mathcal{B}_k(x)\) coincide when \(q=3\), since every face in the \(k\)-th layer is necessarily glued along an edge to a face of \(A_{p,q}(k-1)\). However, when \(q\ge 4\) there will always be new faces only glued to the previous layer at a vertex, and \(\mathcal{B}_k(x)\) will correspond to a sub-animal of \(A_{p,q}(k)\). In, Keller and Peyerimhoff compute the perimeter of these combinatorial balls \(\mathcal{B}_k(x)\). These values are then used to find the *exponential growth* of the graph, defined as, \[\mu= \limsup_{k\to \infty } \frac{\log \text{vol} (\mathcal{B}_k(x)) }{k}.\] They obtain \(\mu= \tau + \sqrt{\tau^2-1}\) for the \(\{p,q\}\)-tessellation, where \(\tau=q-\frac{2}{p-2}\). This value is related to our \(t=(p-2)(q-2)-2\) by \(\tau=\frac{t}{p-2}+2\). Moreover, Theorem [\[thm:layers\]](#thm:layers){reference-type="ref" reference="thm:layers"} can be stated in terms of the combinatorial curvature at a vertex of a \(\{p,q\}\)-tessellation, given by \(\kappa= 1-\frac{q}{2}+\frac{q}{p} = \frac{2-t}{2p}\). In, Mertens and Moore find recursive formulas for the number of vertices and tiles on the sequence \(A_{p,q}(k)\) in order to sample a certain type of random hyperbolic \(\{p,q\}\)-animals that allows them to give computational estimates of hyperbolic percolation thresholds on \(\{p,q\}\)-tessellations. Instead of relying on these independent and related works, we give a complete proof of Theorem [\[thm:layers\]](#thm:layers){reference-type="ref" reference="thm:layers"} to keep the discussion self-contained. In addition, our analysis is aimed at a more detailed description of the perimeter structure of \(A_{p,q}(k)\), which we use to compute the graph parameters of all extremal \(n\)-tile \(\{p,q\}\)-animals. After proving Theorem [\[thm:layers\]](#thm:layers){reference-type="ref" reference="thm:layers"}, we construct and analyze a sequence of spiral animals with \(n\) tiles, of which the complete \(k\)-layered animals are a subsequence. These \(\{p,q\}\)-spirals are denoted by \(S_{p,q}(n)\) and are constructed in roughly the following way: find the maximum \(k\) such that \(\check{n}(k) \leq n\), then attach \(n-\check{n}(k)\) adjacent tiles in one direction along the boundary of \(A_{p,q}(k)\), saturating perimeter vertices one a time along the way. The definition of the sequence \(S_{p,q}(n)\) is made more precise in Definition [\[defn:spiral\]](#defn:spiral){reference-type="ref" reference="defn:spiral"} (see Figure [\[fig:S37-4-5\]](#fig:S37-4-5){reference-type="ref" reference="fig:S37-4-5"}). Before discussing the spirals in more detail, we pause here for a minute to point out the subtleties in the deluge of notation and bookkeeping we employ to keep track of all the parameters we are interested in. In particular, the same set of graph parameters are discussed in three distinct settings, and we use variations on the notation \(v\), \(e\), and \(n\) for each. First, we have the parameters of \(A_{p,q}(k)\), which use the adorned letters presented in Theorem [\[thm:layers\]](#thm:layers){reference-type="ref" reference="thm:layers"} with the layer value as the input. For the spiral \(S_{p,q}(n)\) we use unadorned letters and the tile number as the input, e.g., \(v(n) = v(S_{p,q}(n)), e(n) = e(S_{p,q}(n))\), etc. For a \(\{p,q\}\)-animal \(A\) or it's underlying planar graph, we use unadorned notation and input the animal's name, e.g., \(v(A)\), \(e(A)\), etc. For the spiral \(S_{p,q}(n)\), we compute its parameters in terms of the sequence of vertex degrees on the perimeter of the previous layer of the spiral. To accomplish this, it is useful to label the vertices on the perimeter of \(A_{p,q}(k)\) by \(x_{k,i}\), and to denote by \(d_{k,i}\) the degree of \(x_{k,i}\). The sequences \(d_{k,i}\) are computed recursively at the end of Section 3.
The next theorem states that the spirals in fact attain extremal values in the graph parameters listed above. It is natural to imagine that this should be the case, as the spiral configuration intuitively has the effect of wrapping up as many vertices and edges in the interior as possible, where they can be shared by the most tiles. Thus, interior parameters are maximized while perimeter and overall counts are minimized.
*Spiral-like* animals have also been used by Buchholz and de Launey for exploring the extremal combinatorial problem of edge-minimization for families of \(\{p,q\}\)-animals that do not tessellate the Euclidean plane. They use spiral-like arrangements to obtain asymptotic results for extremal values on the number of internal edges. In that setting, it remains open as to whether these spiral-like arrangements are actually extremal, that is, if they attain and provide the exact minimal optimal values for the number of internal edges. Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"} shows that the minimum perimeter attainable by a \(\{p,q\}\)-animal with \(n\) tiles, which we denote by \(\mathcal{P}_{p,q}(n)\), is equal to \(e_1(n)\). Plugging the Euclidean \(\{p,q\}\) values into Theorem [\[thm:e_1spiral\]](#thm:e_1spiral){reference-type="ref" reference="thm:e_1spiral"}, we recover Harary and Harborth's formulas for extremal values of animals on regular tessellations of the Euclidean plane.
While the Euclidean cases immediately yield closed formulas, in the hyperbolic case it remains open to find a closed formula for \(\mathcal{P}_{p,q}\) in terms of \(n\). Finally, we address the problem of enumerating extremal \(n\)-tile animals for a fixed \(n\), exhibiting several sequences which give unique extremal animals up to isometries. The question of counting extremal animals was first posed by Harary and Harborth in. In the \(\{4,4\}\) case, Kurz proved that squares and pronic rectangles, i.e. polyominoes with \(l^2\) tiles and rectangles with \(l(l+1)\) tiles for \(l\geq 1\), respectively, are unique extremal animals up to isometries. Here, we find sequences of \(\{p,q\}\)-animals which generalize this result for all \(\{p,q\}\) pairs in Theorems [\[thm:uniq_layered\]](#thm:uniq_layered){reference-type="ref" reference="thm:uniq_layered"} and [\[thm:pronic\]](#thm:pronic){reference-type="ref" reference="thm:pronic"}. The rest of the paper is structured as follows: In Section 2, we set up notation and discuss graph theoretic formulae we will need later on. In Section 3, we analyze the layered animals \(A_{p,q}(k)\), we deduce formulas for their parameters to prove Theorem [\[thm:layers\]](#thm:layers){reference-type="ref" reference="thm:layers"}, and we write the substitution rules necessary to construct their perimeter sequence of vertex degrees \(d_{k,i}\). We conclude Section 3 with detailed examples that highlight the difference between the Euclidean and hyperbolic cases. Then, in Section 4, we prove Theorem [\[thm:e_1spiral\]](#thm:e_1spiral){reference-type="ref" reference="thm:e_1spiral"} by finding algebraic expressions for the parameters of \(S_{p,q}(n)\) in terms of the sequences \(d_{k,i}\). In Section 5, we provide the proof of Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"} which states that spirals are indeed extremal. In Section 6, we prove Theorems [\[thm:uniq_layered\]](#thm:uniq_layered){reference-type="ref" reference="thm:uniq_layered"} and [\[thm:pronic\]](#thm:pronic){reference-type="ref" reference="thm:pronic"} regarding unique extremal animals. And finally, in Section 7, we state open problems and give concluding remarks.
# Definitions and Preliminary Results
To begin, we restrict ourselves to animals with no holes, where a hole is defined to be a finite connected component of the complement. We then extend results on extremality to animals with holes as the last piece in proving Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"}. Consider a \(\{p,q\}\)-animal \(A\) with no holes. The edges and vertices of the \(p\)-gons of \(A\) define a planar graph in which all vertices have degree at least 2 and at most \(q\), and its bounded faces are all regular \(p\)-gons. Here, and throughout this paper we use the term faces only to refer to the *bounded* faces of a planar graph. So the number of tiles of \(A\) is equal to the number of faces of its underlying planar graph, which we also refer to simply as \(A\), conflating the notation for convenience. There is also a dual graph \(G'\) associated with any planar graph \(G\), constructed by placing a vertex at each face and connecting two vertices if and only if their corresponding faces share an edge. For an animal \(A\), the dual graph \(A'\) may not be an animal, as some edges may not be contained in any faces. Thus we define a larger class of planar graphs to work in, which contains all graphs defined by animals with no holes.
Then a \(\{p,q\}\)-animal with no holes is necessarily a \(\{p,q\}\)-graph. However, not every \(\{p,q\}\)-graph is the graph of an animal or even a subgraph of the regular \(\{p,q\}\)-tessellation. For instance, we can build a rooted \(\{p,q\}\)-tree starting with a single vertex, the root \(R\), which has \(q\) adjacent vertices. Then we give each of \(R\)'s neighbors \(q-1\) distinct additional neighbors, and then each of those new vertices get \(q-1\) distinct additional neighbors, and so on. After more than \(p/2\) iterations of this, such a graph can no longer live in the \(\{p,q\}\)-tessellation, as these strings starting from \(R\) would necessarily form \(p\)-gons. For the rest of this paper, we always assume that \((p-2)(q-2)\ge4\) for the pair \(\{p,q\}\). We denote the total number of vertices, edges and faces of a fixed graph \(G\) in the same manner as for a \(\{p,q\}\)-animal, using \(v(G)\), \(e(G)\) and \(n(G)\), respectively. Furthermore, we partition the vertices and edges by the number of incident edges and faces, respectively: \[\begin{aligned}
v_i(G) &= \# \text{ vertices of degree } i, \text{ for } 1\le i \le q-1,\\ v_q(G) &= \# \text{ vertices of degree \(q\) incident to only \(q-1\) faces,}\\ v_{int}(G) &= \# \text{ interior vertices, i.e., vertices of degree \(q\) incident to \(q\) faces,}\\ e_i(G) &= \# \text{ edges incident to \(i\) faces, for } i = 0,1,2. \end{aligned}\] All non-interior vertices are called *perimeter* vertices, and these are partitioned by the \(v_i\) according to their degrees. When a fixed \(G\) is understood, we write simply \(v\), \(e\), \(n\), etc, and for its dual graph \(G'\) we write \(v'\), \(e'\), \(n'\), etc. It is easily seen that when \(G\) is a \(\{p,q\}\)-graph, \(p\) and \(q\) switch roles for \(G'\), which is a \(\{q,p\}\)-graph. In general, \(v'\), \(e'\) and \(n'\) can be counted directly by various parameters of \(G\). We know by definition that \(v' = n\); the edges of \(G'\) correspond to edges of \(G\) which are incident to two faces of \(G\), and hence \(e' = e_2\); and a face of \(G'\) corresponds to an interior vertex of \(G\), so \(n' = v_{int}\). We further partition the edges of the dual graph and count its connected components: \[\begin{aligned}
e_{2,i}(G) &= \# \text{ edges in \(G'\) incident to \(i\) faces, for } i = 0,1,2,\\[4pt] c'(G) &= \# \text{ connected components of \(G'\).}\\ \end{aligned}\] When \(G\) is an animal, \(c'\) is always 1 by definition. But here we will obtain more general bounds for all \(\{p,q\}\)-graphs. Using Euler's formula for \(G'\), and also counting its edges with respect to incident faces, we have that: \[\begin{aligned}
n-e_2+v_{int}&=c', \\ q\cdot v_{int} &= e_{2,1}+2e_{2,2} =e_2-e_{2,0}+e_{2,2}. \end{aligned}\] Combining these equations to eliminate \(v_{int}\) yields \[\begin{aligned}
e_2 = \frac{q(n-c')-e_{2,0} +e_{2,2}}{q-1}. \label{eqn:interior-edges} \end{aligned}\] This formula will allow us to use induction to prove that \(e_2(G)\leq e_2(S_{p,q}(n))\) for any \(\{p,q\}\)-graph with exactly \(n\) faces by translating the inequality to the dual graphs. We will also need the following lemma which asserts that the number of tiles decreases when we perform the dual operation. Observe that if a \(\{p,q\}\)-graph is a *tree*, that is a connected graph with no faces, then \(v_{int}=0\) and the dual graph of that component is the empty graph. Thus, it is sufficient to only consider components with faces.
# Complete \(k\)-layered \(\{p,q\}\)-animals {#section:layers}
In this section, we define and study the main properties of the complete \(k\)-layered \(\{p,q\}\)-animals \(A_{p,q}(k)\). Recall from Definition [\[defn:completelayers\]](#defn:completelayers){reference-type="ref" reference="defn:completelayers"} that \(A_{p,q}(1)\) is the regular \(p\)-gon, and then \(A_{p,q}(k)\) is constructed inductively by attaching all allowable tiles to the perimeter of \(A_{p,q}(k-1)\). The graph parameters of \(A_{p,q}(k)\) are denoted by \(\check{v}(k)\), \(\check{e}(k)\), \(\check{n}(k)\), etc. As each new layer is added, these parameters can be tracked by characterizing the geometry of how individual tiles are attached. When adding a tile \(T\) to an animal \(A\), we use the notation \(\deg_{A}(T)\) to denote the number of edges \(T\) shares with \(A\). This is, in fact, the degree of the vertex representing \(T\) in the dual graph of \(A\cup T\). More informally, we describe a tile being added to a given animal as being *\(\epsilon\)-glued* if \(\deg_{A}(T)=\epsilon\). In constructing \(A_{p,q}(k)\) from \(A_{p,q}(k-1)\), we consider each tile individually and examine \(\deg_{A_{p,q}(k-1)}(T)\) for every tile \(T\) in the \(k\)-th layer. We will use this parameter to classify a layer's perimeter vertices by degree, and then find recursive formulas counting the number of each degree in a layer. One key observation is that the intersection of a tile in the \(k\)-th layer with \(A_{p,q}(k-1)\) is a connected path, which may be only a single vertex, in which case the tile is *0-glued*. Then a given perimeter edge gets covered with a single tile, and we must consider if that tile can cover any adjacent edges on the perimeter as well. For instance, a tile that is 2-glued will completely cover one shared vertex with \(A_{p,q}(k-1)\), which is possible if and only if that vertex already has degree \(q\) in \(A_{p,q}(k-1)\). Otherwise 0-glued tiles must be filled in next to a 1-glued tile until the degree of the common perimeter vertex is \(q\) (see Figure [\[fig:A3-recursion\]](#fig:A3-recursion){reference-type="ref" reference="fig:A3-recursion"}).
Inherent in the proof of the recursions in Proposition [\[prop:recurrence\]](#prop:recurrence){reference-type="ref" reference="prop:recurrence"} below is an inductive argument confirming these remarks, which are inextricably linked with each other. For instance, assuming Remark [\[rmk:vdeg\]](#rmk:vdeg){reference-type="ref" reference="rmk:vdeg"}, if \(q\ge4\) there can never be perimeter vertices of degree \(q\), because degree 4 vertices only exist for \(p=3\), and \(p=3\) requires \(q\ge6\). Consequently, there cannot be \(\epsilon\)-glued tiles for \(\epsilon\ge 2\) when \(q\ge 4\). Conversely, if \(p=3\), then \(q\ge6\) and every tile is 0-glued or 1-glued, and the ensuing geometry can be used to show that all perimeter vertices have degree 3 or 4.
Here, we collect the formulas for \(A_{p,q}(k)\) in the hyperbolic case, \((p-2)(q-2)>4\). \[\label{eq:allAparams} \begin{split} \check{n}(k)&= 1 + \frac{p(q-2)}{(\alpha-1)\sqrt{t^2-4}} \left( \alpha^k + \frac{1}{\alpha^{k-1} }-\alpha-1 \right) \\ \check{v}(k)&= \frac{p}{t-2} \left( \alpha^k+ \frac{1}{\alpha^k}-2\right) \\ \check{e}(k) &= \frac{p}{(\alpha-1)\sqrt{t^2-4}} \left( (\alpha+q-1)\alpha^k+\frac{\alpha q-\alpha+1}{\alpha^k}-q(\alpha+1) \right) \\ \check{e}_1(k) &= \frac{p(\alpha+1)}{\sqrt{t^2-4}} \left( \alpha^{k-1}-\frac{1}{\alpha^k} \right) \\ \check{e}_2(k)&= \frac{p}{(\alpha-1) \sqrt{t^2-4}} \left( (\alpha q-\alpha+1) \alpha^{k-1} + \frac{q-1+\alpha}{\alpha^{k-1}}-q(\alpha+1) \right) \\ \check{v}_{int}(k) & = \frac{p}{t-2} \left( \alpha^{k-1}+\frac{1}{\alpha^{k-1}}-2\right) \end{split}\] Now, we establish how to construct the sequences \(d_{k,i}\). To begin with, we order the vertices on the boundary of \(A_{p,q}(k)\), labeling them \(x_{k,i}\) for \(1\le i\le \check{e}_1(k)\).
Now, recall from the proof of Proposition [\[prop:recurrence\]](#prop:recurrence){reference-type="ref" reference="prop:recurrence"} that each vertex in \(A_{p,q}(k)\) generates vertices on the next layer according to its degree. Let \(d_k\) denote the sequence \(\{d_{k,i}\}\). We can compute the sequence \(d_{k+1}\) from \(d_k\) in the following manner. See the ensuing figures. Firstly, \(d_{1}\) is the sequence \(2,2,\dots, 2\) containing \(p\) elements. Then, \(d_{k}\) is constructed from \(d_{k-1}\) by replacing the \(i\)th element \(d_{k-1,i}\) with the string of the degrees of the perimeter vertices of \(A_{p,q}(k)\) associated to \(x_{k-1,i}\). More precisely: If \(p\geq 4\) and \(q\geq 4\), \(d_{k-1,i}\) is replaced by \[\underbrace{ \underbrace{3,2,\dots,2}_{p-3}, \underbrace{3,2,\dots,2}_{p-2}, \dots, \underbrace{3,2,\dots,2}_{p-2} }_{q-d_{k-1,i} \text{ groups of } 3,2,\dots,2}.\] If \(p=3\), and \(q\geq 6\), \(d_{k-1,i}\) is replaced by \[\underbrace{4,3,\dots, 3}_{q-d_{k-1,i}-1}.\] Finally, if \(q=3\), and \(p\geq 6\), \(d_{k-1,i}=2\) is replaced by \[\underbrace{3,2,\dots ,2}_{p-3} \text{ or } \underbrace{3,2,\dots ,2}_{p-4}\] depending whether \(d_{k-1,i}=2\) is preceded by a 2 or a 3, respectively. On the other hand, \(d_{k-1,i}=3\) is simply erased.
# \(\{p,q\}\)-spiral animals
In this section, we turn our attention to \(\{p,q\}\)-spiral animals, counting their graph parameters, and proving Theorem [\[thm:e_1spiral\]](#thm:e_1spiral){reference-type="ref" reference="thm:e_1spiral"} The \(\{p,q\}\)-spiral with \(n\) tiles will be denoted by \(S_{p,q}(n)\). Recall that we refer to its graph parameters as \(e(n), v(n), e_1(n)\), etc. The following procedure for constructing \(S_{p,q}(n)\) gives the spiral in the counter-clockwise direction, as seen in Figure [\[fig:S37-4-5\]](#fig:S37-4-5){reference-type="ref" reference="fig:S37-4-5"}.
For the rest of this section, we always assume that \(\check{n}(k) < n \le \check{n}(k+1)\), and thus \(T_n\) is a tile being added in the \((k+1)\)-th layer. Similar to \(A_{p,q}(k)\), we use the gluing parameter of each subsequent tile to chart the jumps in the sequences \(e_1(n)\) and \(e_2(n)\). Here, \(T_n\) is added to \(S_{p,q}(n-1)\), and so we say \(T_n\) is \(\epsilon\)-glued if \(\deg_{S_{p,q}(n-1)}(T_n)=\epsilon\), with \(\epsilon\ge 1\) since \(T_n\) always shares an edge with \(T_{n-1}\).
This is a simple observation, as when \(T_n\) is \(\epsilon\)-glued it adds \(p-\epsilon\) perimeter edges while covering \(\epsilon\) edges of \(S_{p,q}(n-1)\), and each covered edge is added to the count for \(e_2\). With \(T_n\) in the \((k+1)\)-th layer, note that \(\deg_{S_{p,q}(n-1)}(T_n)\) is also easily determined from \(\deg_{A_{p,q}(k)}(T_n)\), differing by either 0, 1, or 2 additional edges. The first tile \(T_{\check{n}(k)+1}\) has no additional shared edges with other tiles in the \((k+1)\)-th layer, and is 1-glued for \(q\geq 4\) and 2-glued for \(q=3\) (the latter for \(k\geq2\)). Meanwhile, the last tile \(T_{\check{n}(k+1)}\) has two additional shared edges; one with \(T_{\check{n}(k+1)-1}\) and one with the first tile in this layer, \(T_{\check{n}(k)+1}\). Thus, it is 2-glued for \(q\geq 4\) or 3-glued for \(q=3\). For any other value of \(n\), \(T_n\) only shares one additional edge with \(T_{n-1}\), and so \(\deg_{S_{p,q}(n-1)}(T_n)=\deg_{A_{p,q}(k)}(T_n)+1\). Combining this with the analysis in Proposition [\[prop:recurrence\]](#prop:recurrence){reference-type="ref" reference="prop:recurrence"} and Remarks [\[rmk:vdeg\]](#rmk:vdeg){reference-type="ref" reference="rmk:vdeg"} and [\[rmk:e-gluing\]](#rmk:e-gluing){reference-type="ref" reference="rmk:e-gluing"}, we can then determine the exact values of these jumps.
We now turn our attention to the properties of the dual graph of \(S_{p,q}(n)\). Recall that \(n'\) is the number of faces in the dual graph, and we use \(T_{n'}\) and \(T_{n'+1}\) to denote the \(n'\)-th and \((n'+1)\)-th tiles of the dual \(\{q,p\}\)-spiral. Additionally, recall that \(e_{2,i}\) is the number of edges that are incident to \(i\) faces in the dual graph, with \(e_{2,0}\) tracking how far off \(S'_{p,q}(n)\) is from being exactly a \(\{q,p\}\)-spiral.
# Proof of Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"}
We now work towards proving Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"} and the extremality of the graph properties of \(S_{p,q}(n)\). Recall that an animal \(A\) with \(n\) tiles is said to be *extremal* if \(e_2(A)\) and \(v_{int}(A)\) are maximal, and \(e(A), e_1(A)\) or \(v(A)\) are all minimal in the set of animals with \(n\) tiles. First, we show the maximality of \(e_2(n)\) within the class of \(\{p,q\}\)-graphs with \(n\) faces. This requires the following lemma, which asserts that if this holds for the dual structure, then it can be lifted to the original graph.
We use induction and the bounds from Lemma [\[lem:v_intbound\]](#lem:v_intbound){reference-type="ref" reference="lem:v_intbound"} to show that this lifting can always be done, and \(e_2(n)\) is in fact always maximal in the class of \(\{p,q\}\)-graphs.
The maximality of \(e_2(n)\) can immediately be churned into the full litany of extremal properties of spiral animals in the class of \(\{p,q\}\)-graphs. As a corollary, we obtain a proof of the statement of Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"} in the particular case for animals with no holes.
To prove Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"}, it remains to show that this corollary also holds for animals with holes. We will use induction on the number of holes to complete the proof, but first we require the following lemma stating that the number of interior edges in \(S_{p,q}(n)\) plus the total number of edges in \(S_{p,q}(l)\) is at least the number of interior edges in \(S_{p,q}(n+l)\).
Given the extremality of all the pertinent graph parameters, we now briefly show how to derive the equations from for the Euclidean cases.
# Enumeration of extremal \(\{p,q\}\)-animals
In 2008, Kurz enumerated polyominoes, i.e. \(\{4,4\}\)-animals, that attain minimum perimeter for any fixed number of tiles. Moreover, he proved that for all \(l\geq 1\) there exists a unique polyomino with \(l^2\) tiles which is extremal, and also a unique polyomino with \(l(l+1)\) tiles. The former are the square shape polyominoes with side length \(l\), and the latter are pronic rectangles with side lengths \(l\) and \(l+1\). In the general \(\{p,q\}\) setting, we find that the sequence \(A_{p,q}(k)\) corresponds precisely to squares with odd side lengths, and we define three more sequences corresponding to even squares and the two types of pronic rectangles distinguished by the parity of the shorter side. We are then able to generalize Kurz's result and show that each of these sequences yields unique extremal \(\{p,q\}\)-animals for any \(p\) and \(q\). First we define the sequence \(B_{p,q}(k)\) analogously to \(A_{p,q}(k)\); but start with \(B_{p,q}(1)=x_0\), a single vertex. Then \(B_{p,q}(2)\) is the set of \(q\) tiles incident to \(x_0\), and \(B_{p,q}(3)\) is the set of all tiles incident to the perimeter of \(B_{p,q}(2)\), and so on. These form another distinct subsequence of spirals that are related to the \(A_{p,q}(k)\) through the dual graph. Observe that the \(B_{p,q}(k)\) sequence corresponds directly to the square shaped polyominoes with even side length in the \(\{4,4\}\) case.
We will use the duality of these two structures to build an inductive proof of their uniqueness, but first, we require a short proposition establishing a condition under which, up to isometries, an animal and its dual graph are in 1-to-1 correspondence.
To extend now the pronic case, we define two more sequences in a similar fashion to the \(A_{p,q}(k)\) and \(B_{p,q}(k)\). Let \(C_{p,q}(1)\) be the \(\{p,q\}\)-animal which has two tiles glued along a single adjoining edge, and let \(D_{p,q}(1)\) be the \(\{p,q\}\)-graph which is simply a single edge connecting two vertices. Then the sequences \(C_{p,q}(k)\) and \(D_{p,q}(k)\) are defined analogously, with all possible tiles incident to the perimeter added in each subsequent step.
As shown in Figure [\[fig:extremal\]](#fig:extremal){reference-type="ref" reference="fig:extremal"}, it is not always the case that all \(\{p,q\}\)-spirals are *unique* extremal animals. In the \(\{4,5\}\)-tessellation, for example, it is straightforward to construct an animal with \(9\) tiles which is extremal, but not equivalent to \(S_{4,5}(9)\), up to isometries. Note that in this setting \(A_{p,q}(2)\) has 13 tiles, and tiles which are incident to only a vertex of \(A_{p,q}(1)\) naturally appear in pairs in the corners. Then let \(A\) be an animal obtained by removing two of these corner pairs. It is simple to compute that both \(A\) and \(S_{4,5}(9)\) will have \(e_1=16\), and since they also have the same number of tiles, all other graph parameters will also be equal. Hence \(A\) is also extremal. Depicted in Figure [\[fig:extremal\]](#fig:extremal){reference-type="ref" reference="fig:extremal"}, we see that there are exactly five distinct extremal \(\{p,q\}\)-animals with 9 tiles up to isometries, where the image on the far left is \(S_{4,5}(9)\).
# Conclusions and open problems
Recall that Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"} asserts that \(\mathcal{P}_{p,q}(n)\), the minimum perimeter that an \(n\)-tile \(\{p,q\}\)-animal can attain, is equal to \(e_1(n)\). Although here we give an algorithm to iteratively find the values of \(\mathcal{P}_{p,q}(n)\), it remains as an open problem to find closed formulas for \(\mathcal{P}_{p,q}(n)\) in the hyperbolic cases. The substitution rules seen in Section [3](#section:layers){reference-type="ref" reference="section:layers"} hint at the theory of Sturmian words as a promising approach to finding such formulas. Nevertheless, we can derive some interesting properties of \(\mathcal{P}_{p,q}\) as a direct consequence of Corollary [\[cor:e_2bounds\]](#cor:e_2bounds){reference-type="ref" reference="cor:e_2bounds"} and Theorem [\[thm:extremalspiral\]](#thm:extremalspiral){reference-type="ref" reference="thm:extremalspiral"}. For instance, we have that \(\mathcal{P}_{3,q}(n)-\mathcal{P}_{3,q}(n) = \pm 1\) for \(q\ge 6\), \(\mathcal{P}_{p,q}(n)\) is non-decreasing when \(p=4\) or \(\{p,q\}=\{6,3\}\), and \(\mathcal{P}_{p,q}(n)\) is strictly increasing otherwise. These properties are summarized in the following Corollary.
In the hyperbolic case, the minimum perimeter is of the order of the number of tiles. Bounds for \(\mathcal{P}_{p,q}(n)\) for sufficiently large \(n\) can be obtained from the following corollary to Theorem [\[thm:e_1spiral\]](#thm:e_1spiral){reference-type="ref" reference="thm:e_1spiral"} bounding \(e_2(n)\).
We further conjecture a sharp limit for this ratio, based on the ratio for the \(A_{p,q}(k)\) sequence.
Recall that \(e_2(n)\) increments by the gluing parameter of each tile. So when \(q=3\), all tiles are either 2-glued or 3-glued, and thus \[\dfrac{e_2(n)}{n}=\dfrac{2n+(\# \text{ of 3-glued tiles})}{n} = 2+\dfrac{\# \text{ of 3-glued tiles}}{n}.\] And similarly when \(q\ge 4\) all tiles are either 1-glued or 2-glued, and the ratio is \[\frac{e_2(n)}{n}=1+\dfrac{\# \text{ of 2-glued tiles}}{n}.\] So this limit is capturing the ratio of these maximally glued tiles in each case. If this limit holds, then all such ratios for the graph parameters of \(S_{p,q}(n)\) will also have limits given by the corresponding subsequences for \(A_{p,q}(k)\). One application of finding exact formulas for \(\mathcal{P}_{p,q}(n)\) is to explore extremal topological animals. Euclidean animals with minimum perimeter have provided the right shape to produce animals with maximally many holes. We expect that the same will hold true for hyperbolic animals. Represent the maximum number of holes that a \(\{p,q\}\)-animal with \(n\) tiles can have by \(f_{p,q}(n)\). Then, using techniques introduced in, the following topological isoperimetric inequality holds \[\begin{aligned}
f_{p,q}(n)\leq \frac{ n p-2(n-1)-\mathcal{P}_{p,q}(n + f_{p,q}(n))}{p}. \end{aligned}\] This also provides a measure of certain geometric efficiencies, which are optimized when \(f_{p,q}(n)\) attains this upper bound. In the Euclidean case, much of the analysis in solving the question of how many holes an \(n\)-tile animal can have stemmed from understanding these efficiencies, and analyzing the gap between \(f_{p,q}(n)\) and this bound. Finally, it is also natural to ask about algebraic characterizations of combinatorial parameters of higher dimensional extremal animals. For example, one can consider face-minimization properties of structures that are built by gluing together \(d\)-dimensional polytopes along codimension-1 faces in \(\mathbb{R}^d\) or \(\mathbb{H}^d\). In these settings, the analogs of the complete layered animals are straightforward to construct, but it is not immediately clear, for instance, how to generalize the 2-dimensional spirals that interpolate between them. | {'timestamp': '2021-09-14T02:15:21', 'yymm': '2109', 'arxiv_id': '2109.05331', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05331'} |
null | null |
# Introduction
The HAWC, Tibet AS\(_\gamma\) and LHAASO collaborations have reported detection of multiple Galactic ultra-high energy gamma-ray (UHE: \(E_\gamma\geq 100\) TeV) sources. The latest published data from the partially completed LHAASO observatory reveal 12 sources with significant (\(> 7 \sigma\)) detections above several 100 TeV. The completed LHAASO facility will greatly enhance our capability to observe the Galaxy's most extreme particle accelerators. The unprecedented sensitivity in the UHE regime provided by LHAASO and other planned facilities such as CTA and SWGO, is particularly important for efforts to identify Galactic sources of \(>\) PeV cosmic rays (CRs). Acceleration to the knee energy in the CR spectrum is a long-standing issue for the supernova-remnant (SNR) CR origin theory, and though UHE \(\gamma\)-rays provide the best method to identify PeV CR accelerators in our Galaxy, a critical examination of all such sources is warranted. Hard \(\gamma\)-ray spectra at these energies are generally perceived to favour hadronic emission mechanisms, since the Klein-Nishina suppression of inverse Compton (IC) emission at UHEs can, in many circumstances, disfavour hard equilibrium spectra. However, this may not apply if the emission region is moderately photon dominated. Recently, (henceforth '''') have pointed out that the emission from high-power pulsars may naturally account for hard spectra UHE sources. Observations reported by the HAWC collaboration already indicate that UHE \(\gamma\)-ray emission in the vicinity of pulsars with high spin-down powers of \(\dot E > \SI{e36}{\erg\per\second}\) may be common. As most of the reported UHE sources to date are spatially coincident or in close proximity to powerful pulsars, disentangling the contribution from competing sources is a necessary step in current and future investigations. Of the 12 LHAASO detected sources, reported the spectral energy density distributions for a selection of three: LHAASO J2226+6057, LHAASO J1825-1326 and LHAASO J1908+0621. For the latter source, a phenomenological fit comparing a hadronic and a leptonic model was presented. For the leptonic fit, the required power-law index of the electron injection spectrum was −1.75, with a cutoff energy of 800 TeV. Such hard injection spectra are not generally expected, at least within a relativistic shock acceleration scenario. At first sight, this might further favour a hadronic model, however it was assumed that the magnetic energy density in the emission zone exceeded that of the photon field in their model. The other two sources are already listed in the HAWC 100 TeV source catalogue, and consistent matches to the UHE emission using standard theoretical injection spectra were presented in . It is thus important to explore if these results still apply in light of the improved data in the UHE regime, and also if such a scenario might also apply for new sources. It is of course possible, that the situation is more complex, with more than one source powering the detected emission. Naively, one might expect the likelihood that several nearby or overlapping sources accelerate particles to PeV energies is improbable, future high resolution instruments may be necessary to unambiguously determine if this is the case. In this Letter, we revisit the predictions of in light of the latest LHAASO results. We focus on the three sources for which SEDs are available, two of which overlap with HAWC sources previously considered in . In section [2](#sec:Method){reference-type="ref" reference="sec:Method"} we describe the methodology and section [3](#sec:Results){reference-type="ref" reference="sec:Results"} presents the results. At the end, we discuss our findings in section [4](#sec:Conclusion){reference-type="ref" reference="sec:Conclusion"}.
# Method {#sec:Method}
To calculate the emission, we use a simple single-zone treatment. In this model, accelerated electrons are injected over a time frame corresponding to the characteristic age of the associated pulsar into a region with constant homogeneous magnetic field and isotropic radiation fields. The injection rate is assumed to follow a time independent (isotropic) power-law distribution with exponential cut-off: \[\begin{aligned}
Q(E) = \dot{N}_0 \cdot \left(\frac{E}{\SI{1}{\erg}}\right)^{-\alpha} \cdot \exp\left(-\frac{E}{E_{\rm cut}}\right), \end{aligned}\] i.e. we will ignore the evolution of the source, and restrict our attention to the highest energy electrons, for which the cooling time is less than or close to the dynamical age. We note that the impact of considering a finite source age rather than an equilibrium is modest in most cases, but improves the fit and modifies the best fit injection index in the case of younger sources and wide spectral coverage. The total electron spectrum is calculated by taking into account energy losses due to both IC and Synchrotron emission. On timescales shorter than the dynamical time, adiabatic losses are assumed to be negligible. The solution can be obtained using a standard single-zone approach, which has been implemented into the open source code GAMERA. Particle escape is neglected, a valid assumption if the particles diffuse sufficiently slowly, and cool fast. The energy is therefore deposited close to the source and the spectral data points obtained cover the entire emission region. We can check *a posteriori* that these hold. For the radiation fields, we use the large scale Galactic emission model of together with the cosmic microwave background (CMB) at the respective source locations, assuming the sources are located at the position of the pulsar counterparts at their nominal distances. We retain the option to include a local enhancement factor of the diffuse photon field due to nearby photon sources / spiral arms (see ). Between source and detection at Earth, \(\gamma\)-rays are absorbed due to pair production with mostly far infrared (FIR) Galactic radiation fields and the CMB, the latter being important for energies above \\sim 100 TeV. In the worst case, for the source LHAASO J1908+0621 associated to the pulsar PSR 1907+0632, the transmissivity at 1 PeV is 67 %. provide a list of potential particle acceleration sites, including supernova remnants or pulsars, associated to the detected UHE sources. We focus here on the possible contribution from pulsars alone. For LHAASO J2226+6057 there is only one known high-power pulsar in the region of interest, while for the other sources two powerful pulsars are located within \(\SI{1}{\degree}\) from the measured positions. The key physical parameters of these pulsars are provided in table [1](#tab:pulsars){reference-type="ref" reference="tab:pulsars"}, characteristic spin-down age is taken as a proxy for the true age of the pulsar. In each case we then fit the normalisation, the injection index \(\alpha\) and the cutoff energy \(E_{\rm cut}\) (within physically motivated bounds) to the data for a fixed magnetic field for each pulsar association. The errors in determining \(E_{\rm cut}\) are found to be highly asymmetric. We therefore estimate the 95 % confidence interval separately using the probability density function of the \(\chi^2\) distribution, and its dependence on \(E_{\rm cut}\).
# Results {#sec:Results}
We apply the method described above to the sources LHAASO J2226+6057, LHAASO J1825-1326 and LHAASO J1908+0621, for which the spectra are provided in.
#### LHAASO J2226+6057:
This source can be associated to the SNR G106.3+2.7 as well as the pulsar J2229+6114 and its nebula (the ''*Boomerang*'' nebula). The centroid of the \(100\) TeV \(\gamma\)-ray source was also shown by the to be spatially coincident with a molecular cloud, which might favour a hadronic origin. However, given the mature age of the SNR (\(\lesssim 10^4\) years), rather unique conditions are necessary to motivate ongoing PeV acceleration at the SNR shock. We consider here the alternative scenario where the electrons/positrons are accelerated as part of the energy dissipation process of the pulsar wind, for example via shock acceleration. Using the pulsar parameters and photon field associated to the pulsar's position (see Table [1](#tab:pulsars){reference-type="ref" reference="tab:pulsars"}), fits based on the LHAASO data alone are not well constrained. While theoretical predictions for relativistic shock acceleration in general disfavour injection power-law indices \(\alpha < 2\) , equally good fits are possible over a broad range of \(\alpha\), and cut-off energies. Figure [\[fig:J2226\]](#fig:J2226){reference-type="ref" reference="fig:J2226"} compares the LHAASO data with an exemplary model fit for \(B=\SI{3}{\micro\gauss}\) and a fixed injection index \(\alpha=2.2\). The resulting normalisation parameter is \(\dot{N}_0 = \SI{9\pm2e33}{\per\erg\per\second}\) and the cutoff energy is \(E_{\rm cut} = 420_{-130}^{+210}\,\si{\tera\electronvolt}\) at a 95 % confidence level. The cut-off energy is comfortably below the maximum potential drop of the pulsar (see for example or Appendix A). The resulting flux corresponds to a fraction \(\eta = \SI{0.13}{\percent}\) of the pulsar's spin-down power injected into electrons above 1 TeV, or \(\approx\) 1 % for electrons above 1 GeV assuming no break in the injected spectrum. PSR J2229+6114 is therefore easily able to provide the necessary power input. Multi-wavelength models of the source have been explored by other authors,, where their best fits correspond to a \(4~\mu\)G magnetic field, with \(\alpha=2.4-2.5\). Note that in both studies an IR photon field of \(0.3\) eV cm\(^{-3}\) was adopted, which differs from the model of used here (see Table [1](#tab:pulsars){reference-type="ref" reference="tab:pulsars"}). An injection spectrum of \(\alpha=2.2\) is also consistent with the radio spectrum, \(S_\nu \propto \nu^{-0.59}\), measured from the Boomerang nebula. interpreted the break at \(\nu \approx 5\) GHz, as an indication that this is in fact the cooled spectrum, though this would require a magnetic field strength of \(2.6\) mG. Recent work has questioned whether such intense fields can exist on large scales. However, themselves interpret the radio emission as originating in the ''crushed'' part of the nebula, due to reverse shock resulting from the SNR interaction with dense H material in the north-east direction. In this scenario, the termination shock in the south west region would provide more favourable conditions not only for acceleration, but escape into the low density cavity evacuated by the SNR. This would also be consistent with the observed off-set between the emission centres of the Boomerang nebula and the LHAASO emission which is separated by \(\approx\)`<!-- -->`{=html}0.4 ° to the south west. While a detailed numerical model is needed to explore the complex transport expected in such conditions, it is possible to make rough estimates of the expected diffusion coefficient to account for the redial extent of the LHAASO emission. We define a radial diffusion coefficient \(D_{\rm r}= R^2/4 \tau\), where \(R\) is the radius of the observed emission, and \(\tau\) is the minimum of the system age and the total (energy dependent) cooling time. Adopting a half-angle of \(\approx\) 1 ° at a distance of \(0.8\) kpc, we find for example at a representative energy of \(100\)TeV, \(D_{\rm r}=\SI{1.5e27}{\square\centi\meter\per\second}\). While this is smaller than the Galactic average of \(D_{\rm Gal} \approx 10^{26} E_{\rm GeV}^{1/3}~{\rm cm}^2{\rm s}^{-1}\), it is consistent with the Bohm limit for a \(3~\mu\)G field, \(D_{\rm Bohm}(100~{\rm TeV}) = \SI{1.1e27}{\square\centi\meter\per\second}\). LHAASO J2226+6057 is a source of considerable interest with new multi-wavelength data forthcoming. Such broadband coverage will be essential to discriminate different interpretations regarding the origins of the UHE emission.
#### LHAASO J1908+0621:
This source overlaps with the well known VHE source and UHE source eHWC J1907+063. However, since two powerful pulsars are spatially coincident with the LHAASO source, we must consider them separately. We perform a joint fit using the new data from the LHAASO observatory together with the HAWC data. The result for \(B = \SI{3}{\micro\gauss}\) associated to the pulsar PSR 1907+0602 is shown in figure [\[fig:J1908\]](#fig:J1908){reference-type="ref" reference="fig:J1908"} with the orange line. The red data points show the data from LHAASO, and the blue triangles the data from HAWC. We also show the systematic and statistical error band from HAWC as a blue shaded region. The resulting fit parameters for a PSR 1907+0602 association are: \(\dot{N}_0 = \SI{10\pm 2e35}{\per\erg\per\second}\), \(\alpha = 2.66 \pm 0.03\). The best fit cutoff is \(E_{\rm cut} = \SI{10}{\peta\electronvolt}\), but any value above 1.4 PeV is consistent with the data at 95 % confidence level. The Hillas limit for a pulsar with the spoin-down power of PSR 1907+0602 is 4.2 PeV. In figure [\[fig:J1908\]](#fig:J1908){reference-type="ref" reference="fig:J1908"}, this value was chosen for \(E_{\rm cut}\). This scenario requires \(\sim\)`<!-- -->`{=html}40% of the spin-down power of the pulsar in electrons above 1 TeV. To avoid exceeding the available energy budget, electrons should be injected at a minimum energy of 240 GeV. However, if a break in the injection spectrum below 1 TeV exists, this condition can be relaxed. Data in the GeV to TeV regime can help to detect possible spectral breaks, and evidence for such a break exists. For an association with PSR 1907+0632, the values \(\dot{N}_0 = \SI{4.2\pm 0.5e36}{\per\erg\per\second}\) and \(\alpha = 2.79\pm 0.03\) are obtained. The cutoff energy is constrained to be \(>\SI{2}{\peta\electronvolt}\) at 95 % confidence level. In addition to being in tension with the Hillas limit of 1.8 PeV, to account for the emission approximately 7 times available pulsar power must be transferred to electrons above 1 TeV. It is therefore highly unlikely that PSR 1907+0632 alone can account for this source, even allowing for evolution of the injected power over the cooling timescale. We also compare the fit here to the equilibrium model from shown as the blue dashed dotted line in figure [\[fig:J1908\]](#fig:J1908){reference-type="ref" reference="fig:J1908"}. The curve, developed initially to fit just the HAWC data, provides a reasonable match to the ultra high energy emission detected by LHAASO. Both models are nearly identical at low energies, but diverge slightly at higher energies due to the higher cutoff energy in the new model. A more gradual turnover at the highest energies is needed to match the slightly reduced flux values of the LHAASO relative to HAWC data in the overlapping energy range. As shown previously , data from the IRAS survey excludes enhancement in the FIR radiation fields in the range from 60 μm--100 μm and the radiation fields cannot be very different from the average Galactic field, leaving the magnetic field as the only uncertain environmental parameter. The fit parameters for larger magnetic field values result in smaller \(\alpha\) and smaller \(E_{\rm cut}\). Insisting that \(\alpha \gtrsim 2.0\) implies an upper limit on the magnetic field strength of \\sim 8 μ, with a resulting cutoff energy of \(730_{-310}^{+810}\,\si{\tera\electronvolt}\) at 95 % confidence level. However, additional physical effects such as energy dependent confinement could in principle allow injection spectra harder than \(\alpha = 2.0\) to occur, and should not be ruled out *a priori*. For \(B = \SI{3}{\micro\gauss}\) the cutoff energy is constrained to be very close to 1 PeV. In scenarios with larger \(B\)-fields, the value for \(E_{\rm cut}\) is reduced, but also less well constrained. Even for \(B = \SI{8}{\micro\gauss}\), a cutoff energy of 1 PeV is still consistent within errors. A detection of radio emission from the source region can help to constrain the magnetic field, but no emission on the scales of the LHAASO source extend was detected so far. calculated upper limits at 1.5 GHz and 6 GHz based on the non-detection of 1 m/ (10 μ/) for a beamsize of \(\SI{51.1}{\arcsecond}\times\SI{39.5}{\arcsecond}\) (\(\SI{12.2}{\arcsecond}\times\SI{8.6}{\arcsecond}\)). With an estimated radial size of 1 ° these limits translate into upper limits of 3.0 × 10^−13^ /c^2^m/s (2.3 × 10^−13^ /c^2^m/s). The model is not necessarily consistent with these limits, but if electrons are indeed injected only above 240 GeV to fulfil the energy requirement, or a break in the injection spectrum below TeV energies exists, the emission is below the upper limit. Models with lower minimum injection energy and without a break fulfilling the upper limit exist as well: In a different approach, we required consistency with the radio limits and performed a fit with the \(B\)-field as a free parameter, but a fixed value for \(\alpha = 2.2\). This leads to \(B = \SI{5.7\pm 0.3}{\micro\gauss}\), \(\dot{N}_0 = \SI{2.8\pm 0.2 e35}{\per\erg\per\second}\) and \(E_{\rm cut} = 830_{-220}^{+360}\,\si{\tera\electronvolt}\). In this scenario, 30 % of the spin-down power is injected into electrons above 1 TeV and a minimum injection energy of 10 GeV is required to be consistent with energy constraints. The diffusion coefficient for \(B = \SI{3}{\micro\gauss}\) at 100 TeV for a source size of 1 ° is 1.3 × 10^28^ c^2^m/s, which is an order of magnitude above the value for Bohm diffusion (1.1 × 10^27^ c^2^m/s). For \(B = \SI{5.8}{\micro\gauss}\) used in the second model, the diffusion coefficient is 3.9 × 10^28^ c^2^m/s, again far above the corresponding Bohm diffusion coefficient of 5.8 × 10^26^ c^2^m/s.\
#### LHAASO J1825-1326:
The environment of this source is equally complex. As with LHAASO J1908+0621, it also possesses a HAWC counterpart, and a similar joint fit was performed. The magnetic field in the electron cooling zone was again initially fixed to be 3 μ. Two candidate pulsars are potentially associated to LHAASO J1825-1326: PSR J1826-1334 and PSR J1826-1256. For an association to PSR J1826-1334 the fit gives \(\dot{N}_0 = \SI{5\pm 2e35}{\per\erg\per\second}\), \(\alpha = 2.23\pm 0.09\) and \(E_{\rm cut}\) is constrained to \(390_{-150}^{+340}\,\si{\tera\electronvolt}\) at a 95 % confidence level. This model corresponds to \(\eta = \SI{50}{\percent}\) and is shown in Figure [\[fig:J1825\]](#fig:J1825){reference-type="ref" reference="fig:J1825"} with the orange curve. A minimum injection energy of 90 GeV is required to not exceed the available power input from the pulsar, but a spectral break below TeV energies relaxes this condition. The cutoff energy of the model is consistent with the Hillas limit of 4.2 PeV. The \(\gamma\)-ray emission from the fit for an association to PSR J1826-1256 is nearly identical for the energy range shown, the resulting fit parameters in this case are \(\dot{N}_0 = \SI{4\pm 2e35}{\per\erg\per\second}\), \(\alpha = 2.39\pm 0.09\) and \(E_{\rm cut} =410_{-160}^{+380}\,\si{\tera\electronvolt}\) at a 95 % confidence level, with corresponding efficiency \(\eta = \SI{20}{\percent}\). A minimum injection energy of 25 GeV is required if no spectral break below 1 TeV exists. The cutoff energy is below the Hillas limit of 4.7 PeV. The model from for the HAWC source eHWC J1825-134 is depicted in figure [\[fig:J1825\]](#fig:J1825){reference-type="ref" reference="fig:J1825"} with a dashed dotted blue curve and follows closely the model from the joint fit. While the \(B\)-field was the same in both cases, the radiation fields were not: In enhancement factors of 3 and 5 were used for the association to PSR J1826-1334 and PSR J1826-1256, respectively. Due to the presence of a cooling break in the new model, no enhancement is needed. Fits for higher \(B\)-field value decrease the resulting value of \(\alpha\). For example, For \(B = \SI{5}{\micro\gauss}\) and an association to PSR J1826-1334 the resulting parameters are \(\dot{N}_0 = \SI{1.8\pm 0.8e35}{\per\erg\per\second}\), \(\alpha = 1.9\pm 0.1\) and \(E_{\rm cut} = 310_{-110}^{+200}\,\si{\tera\electronvolt}\). An enhancement in the radiation field will increase the resulting value for \(\alpha\). As shown in , enhancements of the IR and UV fields up to a factor of 16 are compatible with upper-limits from IRAS data, and therefore a variety of models for larger \(B\)-field values are viable as long as the source-specific radiation fields are not constrained further. As for LHAASO J1908+0621, no radio emission on the scales of the LHAASO source extend was detected so far. calculated an upper limit at 1.4 GHz based on the non-detection of 0.24 m/ for a beamsize of \(\SI{9.24}{\arcsecond}\times\SI{6.43}{\arcsecond}\). With an estimated radial size of 1 ° given by this limit translates into an upper limit of 2.3 × 10^−12^ /c^2^m/s. The models developed above fulfil this constraint. The diffusion coefficients at 100 TeV are 2.3 × 10^28^ c^2^m/s (PSR J1826-1334) and 6.0 × 10^27^ c^2^m/s (PSR J1826-1256), both above the Bohm diffusion case of 1.1 × 10^27^ c^2^m/s.\
As shown above, reasonable leptonic single-zone models exist for all of the sources. Due to the uncertainty in the environmental conditions, considerable flexibility in the fitting models is permitted. In general however, larger magnetic field values require harder injection spectra, while enhanced radiation fields have the opposite effect. To constrain parameters and to confirm or rule out different scenarios, it is therefore necessary to have either more information about the specific environmental conditions, or to include multi-wavelength data in the modelling [^1]. This requires a detailed source-specific investigation. For each source, multiple possible counterparts exist and taking into account different sources might be necessary, especially at energies below several TeV. In all models, the cutoff energies were constrained to be between few 100 TeV to more than a PeV in some cases. Nearby pulsars can supply sufficient power to account for the observed flux levels, and therefore the nebulae of high spin-down power pulsars may frequently house particles with energies in excess of 100 TeV. The estimated diffusion coefficients are orders of magnitudes below the interstellar CR diffusion coefficients, but similar to estimates for the region around Geminga. Such low diffusion coefficients could be caused by self-confinement.
# Conclusion {#sec:Conclusion}
In this Letter we have shown, that the emission from LHAASO J2226+6057, LHAASO J1825-1326 and LHAASO J1908+0621 is consistent with simple one zone models of inverse Compton emission, provided that, in the case of LHAASO J2226+6057 and LHAASO J1825-1326, the magnetic fields in these systems are somewhat lower than typical Galactic average values, or local radiation field energy densities somewhat higher than the large-scale average. For each source there exists at least one plausibly associated pulsar providing the necessary power to account for the observations above 1 TeV. This fact, coupled with the validity of the IC emission scenario as demonstrated here, can be seen as strong evidence that high spin-down power pulsars routinely accelerate electrons to hundreds of TeV and even PeV energies. The detection of PeV \(\gamma\)-rays from the Crab nebula, recently reported by LHAASO, supports this claim. Since all such neutron stars are thought to launch an ultra-relativistic wind, such extreme accelerators may be surprisingly common, provided they satisfy the associated Hillas condition. However, as discussed in , special conditions may be necessary for these sources to produce detectable fluxes. Regions with strong radiation fields and/or low magnetic-fields exist in many parts of the Galaxy, providing suitable conditions for hard inverse Compton emission at energies above 100 TeV. Radiation dominated sources may be particularly prevalent close to the Galactic centre region where the energy density in the diffuse photon field increases rapidly with decreasing Galacto-centric radius. Future detection of new UHE \(\gamma\)-ray sources are anticipated, with many new results expected from the LHAASO observatory. Future southern hemisphere observatories such as CTA South and SWGO, which are better positioned to observe the Galactic plane and central region, are highly desirable in this effort. The search for a firm identification of the sources of PeV protons and nuclei is ongoing. To confirm or rule out different scenarios, it is crucial to have detailed information about the environmental conditions and to take multiwavelength data into account. In some cases, it may be necessary to consider multi-source models. Very high energy \(\gamma\)-ray spectral information alone is in most cases not sufficient to discriminate between leptonic and hadronic scenarios, but we note that *spatially resolved* VHE-UHE emission might well be, in particular if the energy-dependent morphology implied by the cooling limited IC picture presented here can be demonstrated. The CTA project is particularly important in this regard . | {'timestamp': '2022-01-05T02:00:49', 'yymm': '2109', 'arxiv_id': '2109.05296', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05296'} |
# Introduction {#sec1}
Opacity is a confidentiality property, which characterizes the scenario whether some "secret\" of a system is ambiguous to a potential malicious intruder based on the evolution of the system. In other words, an opaque system always holds the plausible deniability for its "secret\" during its execution. Opacity can be adapted to capture a variety of privacy and security requirements in cyber-physical systems, including event-driven dynamic systems and time-driven dynamic systems . The notion of opacity initially appeared in the computer science literature to analyze cryptographic protocols. Whereafter, various versions of opacity were characterized in the context of discrete-event systems (DESs). For instance, in automata-based models, several notions of opacity have been proposed and studied, which include current-state opacity , initial-state opacity , \(K\)-step opacity , infinite-step opacity , and language-based opacity . Some more efficient algorithms to check them have also been provided in . Furthermore, when a given system is not opaque, its opacity enforcement problem has been extensively investigated using a variety of techniques, including supervisory control , insertion or edit functions , dynamic observers , etc. Recently, verification and/or enforcement of opacity have been extended to other classes of settings, including Petri nets , stochastic systems , modular systems , networked systems , etc. In recent literature , the authors studied opacity preserving (bi)simulation relations using an abstraction-based technique for nondeterministic transition systems and metric systems, respectively. Some applications of opacity in real-world systems have also been provided in the literature, see, e.g., . In automata-based models, different notions of opacity can capture different security requirements. For example, current-state opacity (CSO) characterizes that an intruder cannot determine for sure whether a system is currently in a secret state. In Location-Based Services (LBS), a user may want to hide his/her initial location or his/her location (e.g., visiting a hospital or bank) at some specific previous instant. Such requirements can be characterized by initial-state opacity (ISO) and \(K\)/infinite-step opacity (\(K\)/Inf-SO)[^1]. However, these four standard versions of opacity have some limitations in practice. Specifically, they cannot capture the situation that an intruder can never infer for sure whether the system has passed through a secret state based on his/her observation. In other words, even though a system is "opaque\" in the standard sense, the intruder may necessarily determine that a secret state must has been passed through. To this end, in , a strong version of opacity called *strong \(K\)-step opacity* is proposed to capture that the visit of a secret state cannot be inferred within \(K\) steps. Also, an algorithm that has complexity \(O(2^{|X|+|X|^2})\) is provided to check it using so-called \(K\)-delay trajectory estimators and R-verifiers, where \(|X|\) is the number of states of a system. Inspired by , in this paper we are interested in defining and verifying two other strong versions of opacity in partially-observed nondeterministic finite-state automata (NFAs), called *strong current-state opacity* (SCSO) and *strong initial-state opacity* (SISO), respectively. They mean that if a run passes through a secret state at the current time (resp., at the initial time), there exists another run that never passes through a secret state and has the same observation as the previous run has. Obviously, they have higher-level confidentiality than the standard versions. Note that, by the definition of strong \(K\)-step opacity proposed in , we know readily that strong \(K\)-step opacity reduces to standard CSO when \(K=0\). Thus, we conclude that SCSO implies the strong \(0\)-step opacity but actually these two strong versions of opacity are incomparable when \(K\geq 1\). Therefore, the proposed notions of SCSO and SISO are clearly different from all the existing ones in the literature. Further, we develop a novel method to verify SCSO and SISO from a span-new perspective. Specifically, the main contributions of this paper are summarized below.
- To overcome the inadequacy that the standard CSO and ISO cannot perfectly characterize high-level privacy requirements, we propose two stronger versions of opacity, called SCSO and SISO. They can characterize higher-level confidentiality than the standard versions.
- For SCSO, we construct an information structure using a novel concurrent-composition technique. And then, we propose a verification criterion of SCSO based on *leaking secret states* defined by us in the proposed structure. This verification approach costs time \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\), where \(|X|\) (resp., \(|\Sigma|, |\Sigma_o|, |\Sigma_{uo}|\)) is the number of states (resp., events, observable events, unobservable events) of a system.
- Regarding the verification problem of SISO, using the novel concurrent-composition technique, we propose another information structure in which the information that whether a system has passed through a secret initial state can be characterized. This approach has also time complexity \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\).
The rest of this paper is arranged as follows. Section [2](#sec2){reference-type="ref" reference="sec2"} provides preliminaries needed in this paper, including the system model, the formal definitions of standard CSO and ISO, and their limitations on confidentiality. Section [3](#sec3){reference-type="ref" reference="sec3"} formalizes the notion of SCSO, and proposes a novel information structure for its verification. In Section [4](#sec4){reference-type="ref" reference="sec4"}, the notion of SISO is proposed, and its verification approach is reported. In Section [5](#sec5){reference-type="ref" reference="sec5"}, we conclude this paper with a brief discussion on how to use the proposed concurrent-composition structure to verify *strong infinite-step opacity*, which can be seen as a strong version of the standard Inf-SO.
# Preliminaries {#sec2}
## System model {#subsec2.1}
A DES of interest is modeled as a *nondeterministic finite-state automaton* (NFA) \(G=(X,\Sigma,\delta, X_0)\), where \(X\) is the finite set of states, \(\Sigma\) is the finite set of events, \(X_0\subseteq X\) is the set of initial states, \(\delta: X\times\Sigma\rightarrow 2^{X}\) is the transition function, which depicts the system dynamics: given states \(x,y\in X\) and an event \(\sigma\in \Sigma\), \(y\in\delta(x,\sigma)\) implies that there exists a transition labeled by \(\sigma\) from \(x\) to \(y\). Note that for \(x\in X\) and \(\sigma\in\Sigma\), if there exists no \(y\in X\) such that \(y\in\delta(x,\sigma)\), then \(\delta(x,\sigma)=\emptyset\). We can extend the transition function to \(\delta :X\times\Sigma^{\ast}\rightarrow 2^{X}\) in the usual manner, where \(\Sigma^{\ast}\) denotes the *Kleene closure* of \(\Sigma\), consisting of all finite sequences composed of the events in \(\Sigma\) (including the empty sequence \(\epsilon\)), see, e.g., for more details on DESs. We use \(\mathcal{L}(G,x)\) to denote the language generated by \(G\) from state \(x\), i.e., \(\mathcal{L}(G,x)=\{s\in\Sigma^{\ast}: \delta(x,s)\neq\emptyset\}\). Therefore, the language generated by \(G\) is \(\mathcal{L}(G)=\cup_{x_0\in X_0}\mathcal{L}(G,x_0)\). For a sequence \(s\in\mathcal{L}(G)\), we denote its prefix closure by \(Pr(s)\), i.e., \(Pr(s)=\{w\in\mathcal{L}(G): (\exists w^\prime\in\Sigma^{\ast})[ww^\prime=s]\}\). Further, for a prefix \(w\in Pr(s)\), we use the notation \(s/w\) to denote the suffix of \(s\) after its prefix \(w\). We without loss of generality assume that \(G\) is accessible, i.e., all its states are reachable from \(X_0\). As usual, we assume that the intruder can only see partially the system's behavior. To this end, \(\Sigma\) is partitioned into the set \(\Sigma_o\) of observable events and the set \(\Sigma_{uo}\) of unobservable events, i.e., \(\Sigma_o\cup\Sigma_{uo}=\Sigma\) and \(\Sigma_o\cap\Sigma_{uo}=\emptyset\). The natural projection \(P:\Sigma^{\ast}\rightarrow \Sigma_o^{\ast}\) is defined recursively by (\(i\)) \(P(\epsilon)=\epsilon\), (\(ii\)) \(P(s\sigma)=P(s)\sigma,\mbox{ if } \sigma\in \Sigma_o\), (\(iii\)) \(P(s\sigma)=P(s),\mbox{ if } \sigma\in \Sigma_{uo}\), where \(s\in\Sigma^\ast\). We extend the natural projection \(P\) to \(\mathcal{L}(G)\) by \(P(\mathcal{L}(G))=\{P(s)\in\Sigma_{o}^{\ast}: s\in\mathcal{L}(G)\}\). We denote the inverse projection by a mapping \(P^{-1}:\Sigma_o^{\ast}\rightarrow 2^{\Sigma^{\ast}}\), i.e., for any \(\alpha\in\Sigma_{o}^{\ast}\), we have \(P^{-1}(\alpha)=\{s\in\Sigma^{\ast}:P(s)=\alpha\}\). To study opacity of \(G=(X,\Sigma,\delta,X_0)\), we assume that \(G\) has a set of secret states, denoted by \(X_{S}\subset X\). We use \(X_{NS}\) to denote the set of non-secret initial states. Let \(s=s_{1}s_{2}\ldots s_{n}\in\Sigma^{\ast}\) and \(x_0\in X_0, x_i\in X\), \(i=1,2,\ldots,n\). If \(x_{k+1}\in\delta(x_k,s_{k+1})\), \(0\leq k\leq n-1\), we call \(x_0\stackrel{s_1}{\rightarrow}x_1\stackrel{s_2}{\rightarrow}x_2\stackrel{s_3}{\rightarrow}\cdots\stackrel{s_n}{\rightarrow}x_n\) a *run* generated by \(G\) from \(x_0\) to \(x_n\) under \(s\). We write \(x_0\stackrel{s}{\rightarrow}x_n\) (resp., \(x_0\stackrel{s}{\rightarrow}\)) when \(x_1,x_2,\ldots,x_{n-1}\) (resp., \(x_1,x_2,\ldots,x_n\)) are not specified. Note that there may be more than one run under a sequence \(s\in\mathcal{L}(G,x_0)\) based on nondeterminism of \(G\). In other words, \(x_0\stackrel{s}{\rightarrow}x_n\) may denote some runs. \(x_0\stackrel{s_1}{\rightarrow}x_1\stackrel{s_2}{\rightarrow}x_2\stackrel{s_3}{\rightarrow}\cdots\stackrel{s_n}{\rightarrow}x_n\) is called *a non-secret run* if \(x_i\in X\backslash X_S\), \(i=0,1,2,\cdots,n\).
## Standard current-state opacity and initial-state opacity {#subsec2.2}
In this subsection, we first recall the formal definitions of standard CSO and ISO in and , respectively. And then, we discuss their limitations for characterizing information security.
Note that, in the notion of standard CSO was defined for deterministic finite-state automata. Obviously, this property can be readily extended to the nondeterministic setting as in Definition [\[de:2.1\]](#de:2.1){reference-type="ref" reference="de:2.1"}.
The phenomena in Examples [\[ex:2.1\]](#ex:2.1){reference-type="ref" reference="ex:2.1"} and [\[ex:2.2\]](#ex:2.2){reference-type="ref" reference="ex:2.2"} indicate that the standard versions of CSO and ISO have some limitations to some extent for characterizing high-level information security of a system. In order to overcome the aforementioned limitations, in this paper we propose two strong versions of opacity, called *strong current-state opacity* (SCSO) and *strong initial-state opacity* (SISO). Note that these two stronger notions of opacity provide high-level confidentiality.
# Strong current-state opacity {#sec3}
In this section, we formally formulate the strong version of CSO for a system. And then, we propose a novel approach to verify it.
## Notion of strong current-state opacity {#subsec3.1}
Note that if \(G\) is a deterministic finite-state automaton, SCSO in Definition [\[de:3.1\]](#de:3.1){reference-type="ref" reference="de:3.1"} can be rephrased as follows: \(G\) is said to be SCSO (w.r.t. \(\Sigma_o\) and \(X_{S}\)) if \[\begin{split} & (\forall x_0\in X_0, \forall s\in\mathcal L(G,x_0): \delta(x_0,s)\in X_S)\\ & (\exists x^\prime_0\in X_0, \exists t\in\mathcal L(G,x^\prime_0))[(P(s)=P(t))\wedge\\ & (\forall \bar{t}\in Pr(t))[\delta(x^\prime_0,\bar{t})\notin X_S]]. \end{split}\] Now we give a physical interpretation of Definition [\[de:3.1\]](#de:3.1){reference-type="ref" reference="de:3.1"}. The notion of SCSO captures that if a sequence \(s\in\mathcal L(G)\) is generated and a secret state is reached, then there exists a sequence \(t\in\mathcal L(G)\) that has the same projection as \(s\) and its execution generates a run that never passes through a secret state. SCSO has a higher-level confidentiality than standard CSO. In other words, SCSO implies standard CSO, but the converse implication does not hold. For instance, the system \(G\) in Example [\[ex:2.1\]](#ex:2.1){reference-type="ref" reference="ex:2.1"} is standard CSO, but not SCSO since after observing the sequence \(aa\) an intruder certainly concludes that \(G\) has passed secret state \(x_4\) or \(x_5\). This results in the following proposition for SCSO and CSO whose straightforward proof is omitted.
In the following Subsection [3.2](#subsec3.2){reference-type="ref" reference="subsec3.2"}, we study the verification of SCSO based on a novel verification approach, which has time complexity \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\).
## Verifying strong current-state opacity {#subsec3.2}
In this subsection, we focus on the verification of SCSO in Definition [\[de:3.1\]](#de:3.1){reference-type="ref" reference="de:3.1"}. In order to obtain the main result, we need to introduce three notions of *non-secret subautomaton*, *observer* and *concurrent composition* for a given system. Given a system \(G=(X,\Sigma,\delta,X_0)\) and a set \(X_S\subset X\) of secret states, we construct a non-secret subautomaton by deleting all its secret states from \(G\), which is denoted by \(G_{dss}\), where "\(dss\)\" stands for the acronym of "*deleting secret states*\". Note that when we delete a secret state \(x_s\in X_S\), all transitions attached to \(x_s\) are also deleted. Formally, \[\label{eq:3} G_{dss}=(X_{dss},\Sigma_{dss},\delta_{dss},X_{dss,0}),\] where
- \(X_{dss}=\{x\in X\backslash X_S: \mbox{ there is a non-secret run } x_0\stackrel{s}{\rightarrow}x\mbox{ for some }x_0\in X_{NS} \mbox{ and } s\in\Sigma^\ast\}\) stands for the set of states;
- \(\Sigma_{dss}=\{\sigma\in\Sigma: \exists x,x^{\prime}\in X_{dss} \mbox{ s.t. } x^{\prime}\in\delta(x,\sigma)\}\) stands for the set of events;
- \(\delta_{dss}: X_{dss}\times\Sigma_{dss}\rightarrow 2^{X_{dss}}\) stands for the transition function, i.e., for all \(x,x^{\prime}\in X_{dss} \mbox{ and } \sigma\in\Sigma_{dss}, x^{\prime}\in \delta_{dss}(x,\sigma) \mbox{ if } x^{\prime}\in \delta(x,\sigma)\);
- \(X_{dss,0}=X_{NS}\) stands for the set of initial states.
Next, we recall the *observer* of \(G_{dss}\) that is defined by \[\label{eq:4} Obs(G_{dss})=(X^{obs}_{dss},\Sigma^{obs}_{dss},\delta^{obs}_{dss},X^{obs}_{dss,0}),\] where \(X^{obs}_{dss}\subseteq 2^{X_{dss}}\backslash \emptyset\) stands for the set of states; \(\Sigma^{obs}_{dss}\) stands for the observable event set \(\Sigma_{dss,o}\) of \(\Sigma_{dss}\); \(\delta^{obs}_{dss}: X^{obs}_{dss}\times\Sigma^{obs}_{dss}\rightarrow X^{obs}_{dss}\) stands for the (partial) deterministic transition function defined as follows: for any \(q\in X^{obs}_{dss}\) and \(\sigma\in\Sigma^{obs}_{dss}\), we have \(\delta^{obs}_{dss}(q,\sigma)=\{x^{\prime}\in X_{dss}: \exists x\in q, \exists w\in\Sigma_{dss,uo}^\ast \mbox{ s.t. }x^{\prime}\in\delta_{dss}(x,\sigma w)\}\) if it is nonempty, where \(\Sigma_{dss,uo}=\Sigma_{dss}\backslash\Sigma_{dss,o}\); \(X^{obs}_{dss,0}=\{x\in X_{dss}: \exists x_0\in X_{dss,0}, \exists w\in\Sigma_{dss,uo}^\ast \mbox{ s.t. }x\in\delta_{dss}(x_0,w)\}\) stands for the (unique) initial state. For brevity, we only consider the accessible part of observer \(Obs(G_{dss})\). Note that the widely-used *observer* \(Obs(G_{dss})\) is an important information structure that can track all possible states of \(G_{dss}\) consistent with the current observation. Specifically, \(Obs(G_{dss})\) captures the observable behavior of \(G_{dss}\), that is, \(\mathcal{L}(Obs(G_{dss}))=P(\mathcal{L}(G_{dss}))\). The time complexity of computing observer \(Obs(G_{dss})\) for \(G_{dss}\) is \(O(|X|^2|\Sigma_o||\Sigma_{uo}|2^{|X|})\). We refer the reader to for more details on the notion of observer.
Now we are ready to introduce a novel information structure called the *concurrent composition* of system \(G\) and observer \(Obs(G_{dss})\), which will be used to verify SCSO in Definition [\[de:3.1\]](#de:3.1){reference-type="ref" reference="de:3.1"}.
For a sequence \(e\in\mathcal{L}(Cc(G,Obs(G_{dss})))\), we utilize the notations \(e(L)\) and \(e(R)\) to denote its left and right components, respectively. A similar notation is employed to the language \(\mathcal{L}(Cc(G,Obs(G_{dss})))\). Further, we use \(P(e)\) to denote \(P(e(L))\) or \(P(e(R))\) since \(P(e(L))=P(e(R))\) for any \(e\in\mathcal{L}(Cc(G,Obs(G_{dss})))\), which depends on the context. Intuitively, \(Cc(G,Obs(G_{dss}))\) characterizes the following two properties: (i) \(\mathcal{L}(Cc(G,\) \(Obs(G_{dss})))(L)=\mathcal{L}(G)\); and (ii) it tracks a sequence \(s\in\mathcal L(G)\) from \(X_0\) and an observation \(\alpha\in\mathcal L(Obs(G_{dss}))\) from \(X^{obs}_{dss,0}\) (if exists) such that \(P(s)=\alpha\).
In order to present the first main result of this paper, we classify the states whose first components belong to \(X_S\) in \(Cc(G,Obs(G_{dss}))\) defined in [\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"} into two types:
- **leaking secret states**: \((x,q)\in X_{cc}\) with \(x\in X_S\wedge q=\emptyset\);
- **non-leaking secret states**: \((x,q)\in X_{cc}\) with \(x\in X_S\wedge q\neq\emptyset\);
The leaking and non-leaking secret states are interpreted physically as follows. If \((x,q)\in X_{cc}\) is a leaking secret state, then, by Definition [\[de:3.2\]](#de:3.2){reference-type="ref" reference="de:3.2"}, in \(G\) there exists \(x_0\in X_0\) and \(s\in\mathcal{L}(G,x_0)\) such that \(x\in\delta(x_0,s)\), and there exists no observation \(\alpha\in\mathcal L(Obs(G_{dss}))\) such that \(P(s)=\alpha\). This means that an intruder certainly concludes that \(G\) has reached/passed through at least one secret state by observing \(P(s)\). If \((x,q)\in X_{cc}\) is a non-leaking secret state, then there exist \(x_0, x^\prime_0\in X_0\), \(s\in\mathcal{L}(G,x_0)\), and \(t\in\mathcal{L}(G,x^\prime_0)\) such that: (i) \(x\in\delta(x_0,s)\); (ii) \(s\) and \(t\) have the same projection, i.e., \(P(s)=P(t)\); (iii) there exists a non-secret run \(x^{\prime}_0\stackrel{t}{\rightarrow}x^{\prime}\) generated by \(G\) from \(x^\prime_0\) to \(x^\prime\) under \(t\). Based on the above preparation, now we are ready to introduce the main result on verification of SCSO for a system. It reveals that a system is SCSO if and only if its concurrent-composition structure does not contain any leaking secret state.
We end this section by analyzing the time complexity of using Theorem [\[th:3.1\]](#th:3.1){reference-type="ref" reference="th:3.1"} to verify SCSO. Since \(G\) has the complexity of \(O(|X|^2|\Sigma|)\) and computing \(Obs(G_{dss})\) costs time \(O(|X|^2|\Sigma_o||\Sigma_{uo}|2^{|X|})\), the time complexity of constructing \(Cc(G,Obs(G_{dss}))\) is \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\). Consequently, the (worst-case) time complexity of using Theorem [\[th:3.1\]](#th:3.1){reference-type="ref" reference="th:3.1"} to verify SCSO is \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\).
# Strong initial-state opacity {#sec4}
In this section, we focus on the formulation of strong ISO and its verification problem for a system.
## Notion of strong initial-state opacity {#subsec4.1}
Intuitively, SISO in Definition [\[de:4.1\]](#de:4.1){reference-type="ref" reference="de:4.1"} captures that for each sequence \(s\in\mathcal L(G)\) starting from any \(x_0\in X_0\cap X_S\), there exists a sequence \(t\in\mathcal L(G)\) starting from some \(x^\prime_0\in X_{NS}\) such that \(P(t)=P(s)\) and \(G\) generates a non-secret run starting from \(x^\prime_0\) under \(t\). Note that, SISO, compared with standard ISO in Definition [\[de:2.2\]](#de:2.2){reference-type="ref" reference="de:2.2"}, has a higher-level confidentiality obviously. In other words, SISO implies standard ISO, but the converse is not true. Taking the system \(G\) shown in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"} for example, by Definitions [\[de:2.2\]](#de:2.2){reference-type="ref" reference="de:2.2"} and [\[de:4.1\]](#de:4.1){reference-type="ref" reference="de:4.1"}, we conclude readily that \(G\) is standard ISO, but not SISO. Next, we illustrate these two strong versions of opacity in Definitions [\[de:3.1\]](#de:3.1){reference-type="ref" reference="de:3.1"} and [\[de:4.1\]](#de:4.1){reference-type="ref" reference="de:4.1"} are incomparable. In other words, SCSO does not imply SISO, and vice versa. For instance, by Definitions [\[de:3.1\]](#de:3.1){reference-type="ref" reference="de:3.1"} and [\[de:4.1\]](#de:4.1){reference-type="ref" reference="de:4.1"}, the system \(G\) depicted in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"} is SCSO, but not SISO. On the other hand, let us consider the system \(G\) shown in Fig. [\[Fig8\]](#Fig8){reference-type="ref" reference="Fig8"}. We know that \(G\) is SISO, but not SCSO.
## Verifying strong initial-state opacity {#subsec4.2}
Inspired by the verification of SCSO in Section [3](#sec3){reference-type="ref" reference="sec3"}, we, in this subsection, use a minor variant of the concurrent composition approach to verify SISO of a system. Consider a system \(G=(X,\Sigma,\delta,X_0)\) in which the set of secret states is denoted by \(X_S \subset X\). \(G_{dss}\) stands for the non-secret subautomaton obtained by deleting all secret states in \(G\), which is defined in Eq. [\[eq:4\]](#eq:4){reference-type="eqref" reference="eq:4"}. \(Obs(G_{dss})\) defined in Eq. [\[eq:5\]](#eq:5){reference-type="eqref" reference="eq:5"} stands for the observer of \(G_{dss}\). To check SISO in Definition [\[de:4.1\]](#de:4.1){reference-type="ref" reference="de:4.1"}, we construct an *initial-secret subautomaton* from \(G\), denoted by \(\hat{G}\), as follows: \[\label{eq:6} \hat{G}=(\hat{X},\hat{\Sigma},\hat{\delta},\hat{X}_0),\] where
- \(\hat{X}=\{x\in X: \exists x_0\in {X_0}\cap {X_S},\exists s\in\Sigma^\ast\mbox{ s.t. }x\in\delta(x_0,s)\}\) stands for the set of states;
- \(\hat{\Sigma}=\{\sigma\in\Sigma:\exists x,x^{\prime}\in\hat{X} \mbox{ s.t. } x^{\prime}\in\delta(x,\sigma)\}\) stands for the set of events;
- \(\hat{\delta}: \hat{X}\times\hat{\Sigma}\rightarrow 2^{\hat{X}}\) stands for the transition function, i.e., for all \(x,x^{\prime}\in\hat{X}\) and \(\sigma\in\hat{\Sigma}\), \(x^{\prime}\in\hat{\delta}(x,\sigma)\) if \(x^{\prime}\in\delta(x,\sigma)\);
- \(\hat{X}_0={X_0}\cap {X_S}\) stands for the set of secret initial states.
Note that when we "delete\" a state that is inaccessible from the set \(\hat{X}_0\) of secret initial states, we also delete all transitions attached to that state. Obviously, \(\hat{G}\) is a subautomaton of \(G\) and is obtained by deleting all its states unreachable from \(\hat{X}_0\). Next, we construct the *concurrent composition* of \(\hat{G}\) and \(Obs(G_{dss})\) for \(G\), which also is an NFA, as follows. \[\label{eq:7} Cc(\hat{G},Obs(G_{dss}))=(\hat{X}_{cc},\hat{\Sigma}_{cc},\hat{\delta}_{cc},\hat{X}_{cc,0}),\] where
- \(\hat{X}_{cc}\subseteq \hat{X}\times 2^{X}\) stands for the set of states;
- \(\hat{\Sigma}_{cc}=\{(\sigma,\sigma):\sigma\in\hat{\Sigma}_o\}\cup\{(\sigma,\epsilon):\sigma\in\hat{\Sigma}_{uo}\}\) stands for the set of events, where \(\hat{\Sigma}_o\) (resp., \(\hat{\Sigma}_{uo}\)) is the set of observable (resp., unobservable) events of \(\hat{\Sigma}\);
- \(\hat{\delta}_{cc}: \hat{X}_{cc}\times\hat{\Sigma}_{cc}\rightarrow 2^{\hat{X}_{cc}}\) is the transition function defined as follows: for any state \((x,q)\in \hat{X}_{cc}\), and for any event \(\sigma\in\hat{\Sigma}\),
- when \(q\neq\emptyset\),\
(a) if \(\sigma\in\hat{\Sigma}_o\), then \[\begin{split} & \hat{\delta}_{cc}((x,q),(\sigma,\sigma))=\{(x^\prime,q^\prime)\in \hat{X}_{cc}: x^\prime\in\hat{\delta}(x,\sigma),\\ & q^\prime=\delta^{obs}_{dss}(q,\sigma) \mbox{ if } \delta^{obs}_{dss}(q,\sigma) \mbox{ is well-defined, } q^\prime=\emptyset \\ & \mbox{otherwise}\}; \end{split}\] (b) if \(\sigma\in\hat{\Sigma}_{uo}\), then \[\hat{\delta}_{cc}((x,q),(\sigma,\epsilon))=\{(x^\prime,q)\in \hat{X}_{cc}: x^\prime\in\hat{\delta}(x,\sigma)\}.\]
- when \(q=\emptyset\),\
(a) if \(\sigma\in\hat{\Sigma}_o\), then \[\hat{\delta}_{cc}((x,\emptyset),(\sigma,\sigma))=\{(x^\prime,\emptyset)\in \hat{X}_{cc}: x^\prime\in\hat{\delta}(x,\sigma)\};\] (b) if \(\sigma\in\hat{\Sigma}_{uo}\), then \[\hat{\delta}_{cc}((x,\emptyset),(\sigma,\epsilon))=\{(x^\prime,\emptyset)\in\hat{X}_{cc}: x^\prime\in\hat{\delta}(x,\sigma)\}.\]
- \(\hat{X}_{cc,0}=\hat{X}_0\times\{X^{obs}_{dss,0}\}\) stands for the set of initial states.
In order to present the verification criterion of SISO in Definition [\[de:4.1\]](#de:4.1){reference-type="ref" reference="de:4.1"}, similarly to the argument in Section [3](#sec3){reference-type="ref" reference="sec3"}, we classify the states in \(Cc(\hat{G},Obs(G_{dss}))\) defined in [\[eq:7\]](#eq:7){reference-type="eqref" reference="eq:7"} into two types:
- **leaking states**: for all \((x,q)\in\hat{X}_{cc}\) with \(x\in\hat{X}\wedge q=\emptyset\);
- **non-leaking states**: for all \((x,q)\in\hat{X}_{cc}\) with \(x\in\hat{X}\wedge q\neq\emptyset\);
We now give an interpretation of leaking and non-leaking states as follows: if \((x,q)\in\hat{X}_{cc}\) is a leaking state, by the construction of \(Cc(\hat{G},Obs(G_{dss}))\), then in \(G\) there exists a secret initial state \(x_0\in X_0\cap X_S\) and a sequence \(s\in\mathcal{L}(G,x_0)\) such that \(x\in\delta(x_0,s)\), and there exists no \(x^{\prime}_0\in X_{NS}\) and \(t\in\mathcal{L}(G,x^{\prime}_0)\) such that \(x^{\prime}_0\stackrel{t}{\rightarrow}\) is a non-secret run and \(P(t)=P(s)\). If \((x,q)\in\hat{X}_{cc}\) is a non-leaking state, then in \(G\) there exists \(x_0\in X_0\cap X_S\), \(x^\prime_0\in X_{NS}\), \(s\in\mathcal{L}(G,x_0)\), and \(t\in\mathcal{L}(G,x^\prime_0)\) such that: (i) \(P(t)=P(s)\), and (ii) \(G\) generates a non-secret run starting from \(x^\prime_0\) under \(t\).
# Concluding remarks {#sec5}
In this paper, we characterized two strong versions of opacity called strong current-state opacity and strong initial-state opacity, respectively. They have higher-level confidentiality than the standard versions. Further, we constructed two information structures using a novel concurrent-composition technique to verify these two strong versions of opacity, which have (worst-case) time complexity \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\). Motivated by and our work, a strong version of standard Inf-SO, called *strong infinite-step opacity* (Inf-SSO), is formally formulate in the context of nondeterministic settings as follows.
By Definition [\[de:5.1\]](#de:5.1){reference-type="ref" reference="de:5.1"}, one sees that \(G\) is Inf-SSO if and only if for all \(x_0\in X_0\) and for all \(s\in\mathcal{L}(G,x_0)\), there exists a non-secret run \(x^\prime_0\stackrel{t}{\rightarrow}\) for some \(x^\prime_0\in X_{NS}\) and some \(t\in\mathcal{L}(G,x^\prime_0)\) such that \(P(t)=P(s)\). We can use the concurrent composition \(Cc(G,Obs(G_{dss}))\) of \(G\) and \(Obs(G_{dss})\) constructed in Definition [\[de:3.2\]](#de:3.2){reference-type="ref" reference="de:3.2"} to present the result on verification of Inf-SSO, which also has time complexity \(O(|X|^4|\Sigma_o||\Sigma_{uo}||\Sigma|2^{|X|})\). We here omit its proof, which is similar to that of Theorem [\[th:3.1\]](#th:3.1){reference-type="ref" reference="th:3.1"}.
In the future, we plan to exploit the proposed concurrent-composition approach to design algorithms for enforcing strong current-state opacity, strong initial-state opacity and strong infinite-step opacity.
[^1]: For convenience, the notion originally named CSO (resp., ISO, \(K\)-SO, and Inf-SO) is categorized as standard CSO (resp., standard ISO, standard \(K\)-SO, and standard Inf-SO) in this paper.
[^2]: For brevity, the terminology "CSO\" (resp., ISO), is used as the acronym of both "current-state opacity\" (resp., "initial-state opacity\") and "current-state opaque\" (resp., "initial-state opaque\"), adapted to the context.
[^3]: The terminology "SCSO\" is the acronym of both "strong current-state opacity\" and "strongly current-state opaque\", which depends on the context.
[^4]: Similar to the terminology "SCSO\", in this paper "SISO\" is the acronym of both "strong initial-state opacity\" and "strongly initial-state opaque\", which also depends on the context.
[^5]: Note that, similar to \(Cc(G,Obs(G_{dss}))\), for each state in \(Cc(\hat{G},Obs(G_{dss}))\), its first and second components are also abbreviated by using their respective subscript set. For instance, \((4,\emptyset)\) is the abbreviation of state \((x_4,\emptyset)\). | {'timestamp': '2021-09-14T02:21:06', 'yymm': '2109', 'arxiv_id': '2109.05475', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05475'} |
# Introduction
Owing to large amounts of conversation data and pre-training models , generation-based chatbots have achieved significant advances and even reach human parity on specific testsets. However, the robustness of the pre-trained model is still low with regard to unseen entities. In practice, users often talk with chatbots about latest news and the recently hot topics, which may not appear in the pre-training or fine-tuning corpus. For instance, "COVID-19" is a new term, which does not appear in the training data of Blender [^1], leading to poor performance when a user mentions "COVID-19". As shown in Figure [\[fig:1a\]](#fig:1a){reference-type="ref" reference="fig:1a"}, given an utterance "I've tested positive for COVID-19", Blender yields a bad response "That's great news" because it misunderstands the utterance by the word "positive", which poses a real challenge for building an accurate and robust generation-based chatbot. Existing methods leverage external knowledge to tackle the problem, where chatbots retrieve relevant knowledge about the entities from a external knowledge base and use the retrieved knowledge to help generating appropriated responses. However, these methods heavily depend on the coverage of the knowledge base and the accuracy of knowledge retrieval, which may fail when the entity is not included by the knowledge base or the retrieved knowledge is inappropriate. As shown in Figure [\[fig:1b\]](#fig:1b){reference-type="ref" reference="fig:1b"}, the knowledge retriever fails to retrieve "COVID-19" from the knowledge base, yielding an incorrect response. According to our statistics, the knowledge retrieval failure is not rare in real practice. Taking Reddit as an example, we collect 407 dialogues over 40 topics on the Trendings panel and find that 24.8% of the topic words are polysemous, indicating the probability of incorrect knowledge retrieval, and 47.9% of topic words are not included by the Wikipedia. To date, there are few studies that have investigated how to build a dialogue generation model within which knowledge may be unavailable during inference. We solve this problem by proposing a knowledge enhanced fine-tuning method, trying to understand semantic information of entities based on the context. For example, given the sentence "*I want to submit a paper to EMNLP*\", a person may not know what "*EMNLP*\" is, but he/she can guess that it should be a conference or a journal, based on the context. Similarly, we aim to enhance the semantic representation of unseen entities by guiding the model to learn the meaning of the words only based on the context information. To achieve this, we take Blender as our backbone model, and propose two auxiliary training objectives (Figure [\[fig:1c\]](#fig:1c){reference-type="ref" reference="fig:1c"}) in fine-tuning, dubbed as **K**nowledge **E**nhanced **Blender** (**KE-Blender**). The first objective is *Interpret Masked Word*, which predicts the word's definition based on the context, where the definition is obtained from a knowledge base. The second is *Hypernym Generation*, which predicts the corresponding hypernym of the word given by WordNet. These two introduced training objectives force the model to learn semantic information from the external knowledge base during training, guessing the meaning of the unseen entity with its context, so as to better understand the input utterance and generate relevant responses during inference. Both training objectives do not require further human labeling, which makes it possible for extending to large-scale pre-training. Results on the Wizard of Wikipedia benchmark show that the proposed model brings performance improvement. The proposed method achieves 14.9 and 18.4 PPL on Wizard Test Unseen in the knowledge available setting and unavailable setting, respectively, which outperforms the Blender baselines (16.3 and 19.9 PPL). To further verify the effectiveness of our method in real-world scenarios, we collect 407 dialogues on the Reddit *Trendings* panel, demonstrating the effectiveness of the proposed method in practice. We release our code and dataset at <https://github.com/Nealcly/KE-Blender>.
# Related Work
## Knowledge Enhanced Pre-training
BAIDU-ERNIE uses entity-level masking and phrase-level masking strategy to enhance knowledge into language model. THU-ERNIE incorporates contextual representations with separate KG embeddings. LUKE proposes an entity-aware self-attention to boost the performance of entity related tasks. SenseBERT uses WordNet to infuse the lexical semantics knowledge into BERT. KnowBERT incorporates knowledge base into BERT using the knowledge attention. TNF accelerates pre-training by taking notes for the rare words. Compared with these methods, which enhances the pre-trained encoder by utilizing named entities or knowledge base, we inject knowledge to improve the generation ability of seq2seq models given the unseen word.
## Knowledge Grounded Dialogue Generation
With advances in deep learning, pre-trained language models have shown promising results in dialogue generation. To equip the models with external knowledge, first show that adding user profile information is able to produce a more consistent and engaging response. propose a Transformer memory network to retrieve knowledge from Wikipedia. use two-step decoding, which first generate a response based on context, and then take the generated response and relative knowledge as input to generate a new response. focus on knowledge selection in dialogue generation by utilizing a sequential latent variable model. further enhance the selection module with the posterior information. use reinforcement learning to optimize knowledge selection with unlabeled data. Different from their work, our KE-Blender does not take knowledge as input, because knowledge is only used to enhance our model during training.
# Method
## Task
Suppose that we have a training set \(\mathbb{D}^S = \{\mathbf{U}^S_i, \mathbf{K}^S_i, \mathbf{R}^S_i\}|_{i=1}^{|L|}\), where \(\mathbf{U}^S_i\), \(\mathbf{K}^S_i\) and \(\mathbf{R}^S_i\) are the dialogue context, the external knowledge retrieved from the knowledge base and the response, respectively. In addition to \(\mathbb{D}^S\), we have a test dataset \(\mathbb{D}^P=\{\mathbf{U}^P,\mathbf{R}^P\}\). Unlike \(\mathbb{D}^S\), \(\mathbb{D}^P\) does not contain external knowledge, because associated background knowledge for unseen word is difficult to obtain in real time during inference. Our goal is to learn a dialogue generation model \(P(\mathbf{R}|\mathbf{U};\theta)\) with the help of \(\mathbb{K}^S\), where \(\theta\) is the parameters of the model. It should be noted that, the dialogue generation model \(P(\mathbf{R}|\mathbf{U};\theta)\) generates the response \(\mathbf{R}\) only based on the input context \(\mathbf{U}\), without using knowledge \(\mathbf{K}\) as input. In the following sections, we will introduce the model structure first, and then show how to leverage the external knowledge \(\mathbf{K}\) to enhance the generation model \(P(\mathbf{R}|\mathbf{U};\theta)\) with our two proposed training objectives.
## Baseline {#sec:baseline}
We consider Blender and **K**nowledge **G**rounded **Blender** (KG-Blender) as our baselines in knowledge available and knowledge unavailable settings, respectively.
#### Blender
Given a dialogue context \(\mathbf{U} =\{\mathbf{u}_1...,\mathbf{u}_{l-1}\}\), we first concatenate \(\mathbf{U}\) as a sequence of sentences \(\mathbf{U} = \{x_1, x_2,\dots, x_T\}\). The response is denoted as \(\mathbf{R} = \{y_1, y_2,\dots, y_{T'}\}\). We train our model on the basis of Blender, which is a standard Seq2Seq Transformer architecture with pre-training. In particular, we feed the dialogue context \(\mathbf{U}\) to the encoder of the transformer, and then we obtain hidden representations of the sentence \[\mathbf{h}^{enc} = \textsc{Transformer\_Encoder}(\mathbf{U}) \label{eq:encoder}\] At the \(t\) th step of the decoder, \(\mathbf{h}^{enc}\) and previous output tokens \(y_{1:t-1}\) are then as inputs, yielding a representation using attention \[\mathbf{h}^{dec}_t = \textsc{Transformer\_Decoder}(\mathbf{h}^{enc}, y_{1:t-1})\] The generative probability distribution of \(y_t\) is given by \[p(y_t|\mathbf{U},y_{1:t-1}) = softmax (\mathbf{W}^o \mathbf{h}^{dec}_t + \mathbf{b}^o)\] where \(\mathbf{W}^o\) and \(\mathbf{b}^o\) are trainable parameters. We use the standard Maximum Likelihood Estimation to optimize the model parameters \(\theta\). Given a training pair \((\mathbf{U},\mathbf{R})\), we minimize: \[\mathcal{L}_{dialogue} =-\sum_{t=1}^{T'} \log p(y_t|\mathbf{U},y_{1:t-1}) \label{eq:dialogue_loss}\] We adopt Blender-90M to initialize our Seq2Seq Transformer model, which has been pre-trained on 1.5B training examples from Reddit 2019.
#### Knowledge Grounded Blender
One intuitive baseline to use knowledge is to take both the context and the knowledge as input. In particular, the concatenation of the context \(\mathbf{U}\) and the associated knowledge \(\mathbf{K}\) is fed to the transformer encoder: \[\begin{split} &\mathbf{h}^{enc'} = \textsc{Transformer\_Encoder}([\mathbf{U};\mathbf{K}]) \\ &\mathbf{h}^{dec'}_t = \textsc{Transformer\_Decoder}(\mathbf{h}^{enc'}, y_{1:t-1}) \\ & p(y_t|\mathbf{U},\mathbf{K},y_{1:t-1}) = softmax (\mathbf{W}^o \mathbf{h}^{dec'}_t + \mathbf{b}^o) \end{split}\] Similar to Eq [\[eq:dialogue_loss\]](#eq:dialogue_loss){reference-type="ref" reference="eq:dialogue_loss"}, given a training pair \((\mathbf{U},\mathbf{K},\mathbf{R})\), the loss function is \[\mathcal{L}_{dial\_know} =-\sum_{t=1}^{T'} \log p(y_t|\mathbf{U},\mathbf{K},y_{1:t-1})\] Note that it is difficult to use KG-Blender directly when knowledge is unavailable, because KG-Blender relies knowledge as input.
## Knowledge Enhanced Blender
To build a robust model for knowledge unavailable setting, we consider adding two auxiliary loss during fine-tuning. People try to understand an unseen word based on the context, even if a dictionary is unavailable. To simulate this behavior, we explicitly guide the Blender model to learn the meaning of words only based on the context information.
#### Interpret Masked Word
The first objective is to ask the model to restore the definition of masked words. We can use different methods to select which words should be masked. For example, we could mask proper nouns in the utterance, or pre-defined topic word for specific dataset[^2]. For example, the input text is *"I submit a paper to the EMNLP"*. "*EMNLP*" is replaced by `[MASK]`, yielding *"I submit a paper to the `[MASK]`"*. The definition retrieved from Wikipedia is "*EMNLP is a leading conference in the area of natural language processing and artificial intelligence* ". Then, the pre-trained model is required to restore the definition by consuming the masked utterance as input. In this way, the model is explicitly guided to understand the background knowledge of the masked word given the context. Formally speaking, given a single utterance \(\mathbf{u}_{l-1} = \{x_1, x_2,\dots, x_T\}\), we assume that \(x_i\) is the topic word in \(\mathbf{u}_{l-1}\), and its corresponding definition is denoted as \(\mathbf{K}_{x_i}=\{k_1,k_2,\dots,k_{|K_{x_i}|}\}\). We use the special token `[MASK]` to replace \(x_i\) yielding \(\mathbf{u}'_{l-1} = \{x_1,\dots,x_{i-1},\)`[MASK]`\(, x_{i+1},\dots,x_T'\}\) as the input of Eq [\[eq:encoder\]](#eq:encoder){reference-type="ref" reference="eq:encoder"} in Section [3.2](#sec:baseline){reference-type="ref" reference="sec:baseline"}. To distinguish with the original dialogue generation task, we use a specific start token `[DEFI]` to mark that the target sequence is the definition. Given a training pair \((\mathbf{u}'_{l-1},\mathbf{K}_{x_i})\), the training objective of the interpret masked word is: \[\mathcal{L}_{interpret} =-\sum_{t=1}^{|K_{x_i}|} \log p(k_t|\mathbf{u}'_{l-1},k_{1:t-1})\]
#### Hypernym Generation
We also reconstruct the input utterance by replacing the topic words with the corresponding hypernym. Compared with topic words, the semantic field of its hypernym is more general. We use WordNet to construct our training instances. For instance, given an utterance \(\mathbf{u}_{l-1}=\) {*I submit a paper to the EMNLP*}, we use "*conference*" to replace "*EMNLP*", where "*conference*" is the hypernym of "*EMNLP*", yielding the target sequence \(\mathbf{u}''_{l-1}=\) {*I submit a paper to the conference*}. This training objective aims to guide the model to understand the semantic information of unseen words. We use a specific start token `[HYPE]` to mark the target sequence is the hypernym generation. Given a training pair \((\mathbf{u}_{l-1},\mathbf{u}_{l-1}'')\), the training objective of the hypernym generation is: \[\mathcal{L}_{hypernym} =-\sum_{t=1}^{|u''_{l-1}|} \log p(y''_t|u_{l-1},y''_{1:t-1})\]
#### Training
We optimize the dialogue generation loss with the two external loss at the same time: \[\small \mathcal{L} = \sum_{i=1}^{|L|} \mathcal{L}^i_{dialogue} + \sum_{i=1}^{|K|} \mathcal{L}^i_{interpret} + \sum_{i=1}^{|H|} \mathcal{L}^i_{hpernym}\] where \(|K|\) and \(|H|\) represent the number of training instances for *Interpret Masked Word* and *Hypernym Generation*, respectively.
#### Inference
During inference, we only take the context as input if the additional retrieved knowledge is unavailable. Following , we adopt greedy search to select the highest probability token at each time step. We denote our model as **K**nowledge **E**nhanced **Blender** (KE-Blender) for the remaining of this paper.
# Experiments
## Datasets
**Wizard of Wikipedia** is a knowledge aware chit-chat dialogue benchmark, where each instance has an initial topic given by two annotators. The dataset contains 18,430 training dialogues with 1,365 topics, and each topic is linked to a Wikipedia article. Its test set is split into Test Seen and Test Unseen based on whether the topic is appear in the training set. We evaluate our methods with several baselines on both Test Seen and Test Unseen. There are 148,357 training instances in the Wizard of Wikipedia training set. To enhance knowledge into the model, we further construct 65,072 training pairs for interpret masked word based on Wikipeida and 86,612 training pairs for hypernym generation based on WordNet. **Reddit Trendings** is a test set to simulate real-world settings, by crawling users' dialogue from its Trendings panel in 2021. Reddit Trendings panel contains the latest hot topics, and most of them are not included in the external knowledge bases. We first obtain topic words from the Reddit Trendings panel, then crawl the dialogue based on the topic words. We further filter the datasets by selecting out dialogue that includes at least 2 utterances, yielding a dataset which similar to the Wizard setting. Finally, the dataset consists of 407 utterances over 40 trending topics.
## Setup
We implement KE-Blender with `transformers` and choose `blenderbot-90M` as the pre-trained language model. AdamW with a batch size of 128 is used to optimize parameters. The initial learning rate is set as 1e-5, which is halved in each training iteration. We set the maximum input tokens as 512. To ensure that KE-Blender also works well in knowledge available settings, we also create extra training instances by concatenating the context with the associated knowledge as input.
## Baselines
We compare KE-Blender with Blender and KG-Blender, also drawing the following state-of-the-art methods as reference: **Transformer** is a standard transformer model for dialogue generation. It takes the concatenation of context utterances and the associated knowledge as input. **SKT** uses a sequential latent variable model for knowledge selection, and then generates the response based on the context and the selected knowledge. **DRD** is a pre-training model designed for the low-resource dialogue generation, which decomposes the decoder into independent components. **SKT + GPT-2** feeds the knowledge selected by SKT to GPT-2 for dialogue response generation. **SKT + PIPM + KDBTS** uses posterior information to help prior knowledge selection module, and trains the decoder with knowledge distillation. **KnowledGPT** adopts reinforcement learning to optimize the knowledge selection module, which gives state-of-the-art performance on Wizard. **Blender-FT.** Blender is a large-scale dialogue pre-training model. We fine-tune the Blender on Wizard training set without utilizing external knowledge. **KG-Blender.** We fine-tune Blender on the Wizard training set by concatenating the context and the associated knowledge as the input. In the setting where external knowledge is unavailable, only context is used to generate response.
## Metrics
#### Automatic evaluation metrics:
Following and, models are measured using the perplexity of the ground-truth response (PPL) and unigram F1-score (F1). show that BERT can be used to evaluate the generated response. We employ BERT-based evaluation metrics to evaluate whether the generated response is knowledgeable as supplements to PPL and F1. As shown in Table [\[fig:eval_bert\]](#fig:eval_bert){reference-type="ref" reference="fig:eval_bert"}, the dialogue generation model is first required to generate response \(\hat{\mathbf{R}}\) based on the dialogue context \(\mathbf{U}=\{\mathbf{u}_1,\dots,\mathbf{u}_{l-1}\}\). We use the special token `[MASK]` to replace the topic word in the last context utterance \(\mathbf{u}_{l-1}\). Then a masked language model (i.e. BERT-large) is used to predict the masked topic word using the last context utterance \(\mathbf{u}_{l-1}\) and the generated response \(\hat{R}\). The recall\(@k\) for the masked language model is used to measure the knowledge stored in the dialogue generation model. Intuitively, if a dialogue generation model is more knowledgeable, the masked language model is stronger to predict the masked topic word based on the generated response \(\hat{\mathbf{R}}\) and last context utterance \(\mathbf{u}_{l-1}\).
#### Human evaluation metrics:
Manual evaluations are essential for evaluating dialogue generation. We conduct human evaluations to compare KE-Blender with our baseline Blender and KG-Blender by randomly sampling 200 instances from the Wizard Test Unseen. We define three metrics for manual evaluation, including fluency, knowledgeability and coherence. Each aspect is scored into three grades, 0, 1 and 2, representing "bad", "normal" and "good" respectively. Following , we employ three annotators to do a side-by-side human evaluation, and report the Fleiss Kappa to show the agreement among human annotators.
## Results
Table [\[tab:wizard\]](#tab:wizard){reference-type="ref" reference="tab:wizard"} reports automatic results on the Wizard of Wikipedia dataset.
#### Test Seen vs Test Unseen
Compared with Test Seen, all models perform worse on Test Unseen, especially where knowledge is unavailable. For example, the Blender-FT only achieves F1 scores of 16.5 and 12.9 on Test Seen and Test Unseen, respectively. Compared with several baselines, KE-Blender gives the lowest performance gap, suggesting that KE-Blender is more robust when it comes to Test Unseen.
#### w/ Knowledge During Inference
Previous work focuses on how to better leverage knowledge for dialogue generation. Compared with these strong baselines, KE-Blender performs competitively in the knowledge-available setting, even though KE-Blender is designed for knowledge-unavailable setting. Notably, it achieves the best reported PPL on Wizard. The results are consistent with our intuition. Knowledge grounded methods perform well when knowledge is provided during inference, and our method is robust and does not degrade in the w/ knowledge setting.
#### w/o Knowledge During Inference
When external knowledge is unavailable during the inference stage, knowledge grounded methods cannot be directly applied to this setting since it requires knowledge as an input. Hence, we compare KE-Blender with Blender and KG-Blender. As can be seen from Table [\[tab:wizard\]](#tab:wizard){reference-type="ref" reference="tab:wizard"}, our method shows large advantages in all metrics, achieving a 16.7 F1 score. It shows that our external training objectives can help model to generalize better when meet unseen words.
#### BERT Evaluation
Table [1](#tab:wizard-knowledge){reference-type="ref" reference="tab:wizard-knowledge"} shows the recall of masked topic words predicted by a BERT model, where a higher recall score indicates the stronger correlation between the knowledge and the response. Human's response obtains a higher score, which means our evaluation metric is reasonable and there is still a gap between human's reply and machine's reply. Our method gives the best performance in both settings, demonstrating strong performance when knowledge is absent, which shows that our auxiliary training objectives is able to help model to learn a better semantic representation. Surprisingly, it also outperforms simple knowledge grounded methods when knowledge is available.
#### Human Evaluation
Table [2](#tab:wizard-human){reference-type="ref" reference="tab:wizard-human"} compares KE-Blender with baselines using human evaluation. All models are able to produce fluent response due to the power of pre-training. Inference with the retrieved knowledge is particularly helpful for the model to generate a more knowledgeable and coherent response. When knowledge is unavailable, KE-Blender significantly outperforms Blender and KG-Blender (\(p\)\<0.01) measured in both knowledgeable and coherent, also giving highly competitive results with the model using knowledge as input. The value of Fleiss' Kappa exceed 0.59 on all models, showing a high inter-rater agreement among annotators.
#### Low-Knowledge-Resource Setting
To simulate a low-knowledge-resource setting, we start from using the full knowledge in Wizard Test Unseen, and gradually reduce the amount of knowledge by randomly removing some entries. Figure [\[fig:percent\]](#fig:percent){reference-type="ref" reference="fig:percent"} shows the trends when different percentage of knowledge in Test Unseen is removed. As the ablation knowledge increases, the performance of the two methods significantly decreases. The F1 of KG-Blender sharply decreases from 17.8 to 16.3. Compared with KG-Blender, the rate of decrease is much smaller for KE-Blender, which shows the effectiveness of knowledge enhanced fine-tuning.
#### Reddits Trendings
We train KE-Blender and KG-Blender on Wizard training set, and test on Reddits Trendings, taking off-the-shelf Blender as a reference. The results are reported in Table [\[tab:reddit\]](#tab:reddit){reference-type="ref" reference="tab:reddit"}. Note that there is a domain gap between Wizard and Reddits, which leads to a worse performance on Reddits Trendings. KE-Blender achieves 56.6 and 54.7 PPL on w/knowledge and w/o knowledge settings, respectively, outperforming KG-Blender and off-the-shelf Blender in all settings. When invalid knowledge is used as input, KG-Blender achieves 66.8 PPL in w/knowledge setting, which underperforms w/o knowledge setting (56.3 PPL). This shows that the inappropriate knowledge selection has a destructive impact on models. Integrating with valid knowledge, both models are able to generate more informative responses. Furthermore, KE-Blender gets the best performance gap, which confirms that KE-Blender does not rely on external knowledge base, demonstrating the effectiveness of the proposed auxiliary training objectives.
# Analysis
#### Ablation Study
An interesting question is to explore the contribution of the two auxiliary losses in training. The results are shown in Table [3](#tab:ablation){reference-type="ref" reference="tab:ablation"}. We can see that each loss contributes a lot in automatic evaluation, with F1 increasing largely by adding each objective. When combining the two losses, there is still an improvement but marginal, which indicates the two loss may play similar roles for pre-training model.
#### Case Study
Table [\[tab:case1\]](#tab:case1){reference-type="ref" reference="tab:case1"} shows an example of the model responses on the Wizard of Wikipedia Test Unseen. Under the knowledge-available setting, all models generate reasonable responses with the help of relevant knowledge. Both models mention that "Elvis" was in "jailhouse rock" by consulting the external knowledge base. When knowledge is unavailable, Blender-FT gives a non-informative response because it cannot understand the word "Elvis". In contrast, KE-Blender shows superior performance by producing informative and knowledgeable responses, which directly points out that "Elvis" appears in a lot of movies and also in a few TV shows. This case shows that our model can significantly improves response quality when knowledge is absent, while sustain good performance when knowledge is available.
#### Knowledge Mining
Although we add two additional tasks in training, it is unclear how well the model performs in these two tasks. Therefore, we further evaluate whether explicit knowledge can be recovered from our model given the unseen entity. First, we find that the perplexity of the ground-truth Wikipedia knowledge on Test Unseen is only 6.81. As shown in Table [4](#tab:case3){reference-type="ref" reference="tab:case3"}, our model is able to produce reasonable definition based on context information and the pre-trained knowledge, and generate hypernyms for a given word associated with context. These show that rich-knowledge is stored in KE-Blender during knowledge enhanced fine-tuning, which potentially allows us to ground open domain dialogues without external knowledge.
# Conclusion
We presented KE-Blender for better handling response generation based on unseen words, which enables a model to generate knowledgeable response without external knowledge during inference. To explicitly inject the knowledge into the model, we proposed two training objectives, including interpret masked word and hypernym generation. To simulate real-world scenario, we also released a test set on Reddit Trendings. Results on Wizard and Reddit Trendings show that KE-Blender outperforms several state-of-the-art methods and strong baselines in settings both when external knowledge is available and unavailable. | {'timestamp': '2021-09-14T02:21:37', 'yymm': '2109', 'arxiv_id': '2109.05487', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05487'} |
# Introduction
Multilayer reflective coatings are widely used in focusing and imaging X--ray optical elements and systems. Quite often they are simple periodical structures of two alternating materials containing from tens to thousands layers. Such periodic coatings are highly spectrally and angularly selective reflecting X-ray radiation only in a narrow range of wavelengths or grazing angles. However, various applications of the reflective X-ray optics including those in the X-ray astronomy, soft X-ray microscopy and EUV/BEUV lithography would be much better served by the broadband aperiodic reflective multilayer coatings having extended reflectivity bands both in the wavelength and angular domains. This advantage is related to the ability of such optical elements to effectively utilize the radiation coming from broadband X-ray and EUV sources such free electron lasers, which could also significantly improve the spatial resolution and reduced the speckle artifacts. In many applications it is also important to optimize the multilayer coatings so that they have the maximal integral reflectivity in a broad spectral or angular passband rather than have the maximal peak reflectivity at a specific wavelength or grazing angle. It should be noted that for multilayer coatings the requirements of the high peak reflectivity and wide reflectivity band are in natural contradiction to each other. Finding a satisfactory compromise requires some sort of analytical theory, which is capable of solving the inverse problem for non-periodical multilayer coatings. Such a theory can be then used for the designing reflective coatings with an arbitrary dependence of the reflectivity on the wavelength or grazing angle as well as for the maximization of their integral reflectivity. A popular alternative to the analytical approach is the use of various numerical techniques including, for instance, the fastest descent method, genetic algorithms and needle method for the optimization and inverse problem solution. On the other hand, over decades there have been considerable developments in the analytical description of non-periodic X-ray multilayer structures, which have met with varying success leaving a plenty of room for further improvements. In this paper we are building on our previous works and present a theory of X-ray multilayer coatings with the graded period, where "graded" means that the multilayer period significantly changes at a distance much larger than the period itself. There are two reasons to consider the graded multilayers: (i) they permit the accurate theoretical analysis, which can then be used as a starting point for the solution of the inverse problem, e.g. designing multilayers with the pre-specified reflectivity curve; (ii) manufacturing the multilayers with a gradually changing period is in practice a more realistic task as compared to fabrication of the multilayers having abrupt jumps of the period. Based on the developed theory we then solve the inverse problem for the multilayer coatings and propose a method to maximize their integral reflectivity.
# Coupled wave equations
[\[sect_coupled\]]{#sect_coupled label="sect_coupled"} Propagation and reflection of X-rays from a multilayer structure is assumed to obey the macroscopic electrodynamics. In this paper we consider the X-ray waves whose electrical field vector is perpendicular to the plane of incidence (s-polarization). However, all the results remain valid for the p-polarization if the grazing angle \(\theta\) is small (grazing incidence) or close to \(\pi/2\) (normal incidence). The important case when \(\theta\approx\pi/4\) is beyond our consideration. We will also neglect the roughness of interfaces and other distortions of ideally sharp boundaries of layers in a multilayer. Then the problem can be reduced to solution of the scalar Helmholtz equation in the following form \[\frac{d^2E}{dz^2}+k^2\left[\varepsilon(z)-\cos^2\theta\right]E=0, \label{fa1}\] where \(E\) is the electric field, \(z\) is the distance from the top of a multilayer, \(\varepsilon(z)\) is the space distribution of the dielectric permittivity. For the analytical description of a graded multilayer with ideally sharp interfaces we use the following spatial dependence of the dielectric permittivity \[\label{fa} \varepsilon(z)=\tilde{\mu_{0}}+4\tilde{B}f(s),\quad s=2\int\limits_{0}^{z}q(z')dz',\] where \[\tilde{\mu_{0}}=\frac{\varepsilon_{1}+\varepsilon_{2}}{2},\quad \tilde{B}=\frac{\varepsilon_{1}-\varepsilon_{2}}{8},\] \(q(z)\) is the so called inverse period function, \(\varepsilon_{1}, \varepsilon_{2}\) are dielectric constants of the two alternating multilayer materials and \(f(s)\) is an even \(2\pi\) periodical function, which can be written as \[\label{fa5} f(s)=1-2\sum\limits_{n=-\infty}^{\infty}\left[\theta(s-\pi\beta-2\pi n)-\theta(s-\pi(2-\beta)-2\pi n)\right],\] where \(\theta(x)\) is a step-like theta function, \(\beta=d_{1}/(d_{1}+d_{2})\) is a fraction of the period occupied by the more optically dense component 1 and \(d_{1}\) and \(d_{2}\) are the thicknesses of layers 1 and 2, respectively. The dielectric function of a periodical structure can be obtained by assuming \(q(z)=const\) in equation ([\[fa\]](#fa){reference-type="ref" reference="fa"}). The period of the multilayer structure is then \(d=\pi/q\). The Fourier series of \(f(s)\) has the form \[\label{fb} \varepsilon(z)=\mu_{0}+4\sum_{n=1}^{\infty}B_{n}\cos(2n\int\limits_{0}^{z}q(z')dz'),\] where \[\mu_{0}=\eta\varepsilon_{1}+(1-\beta)\varepsilon_{2},\quad B_{n}=\frac{\varepsilon_{1}-\varepsilon_{2}}{2n\pi}\sin\pi n\beta.\] Substituting [\[fb\]](#fb){reference-type="eqref" reference="fb"} into the Helmholtz equation [\[fa1\]](#fa1){reference-type="eqref" reference="fa1"} we obtain \[\label{fc} E''+k^{2}\left[\mu+4B\cos(2\int\limits_{0}^{z}q(z')dz')\right]E=0,\] where \(B=B_{1}\), \(\mu=\mu_{0}-\cos^{2}\theta\). In [\[fc\]](#fc){reference-type="eqref" reference="fc"} following the papers we kept only the first spatial harmonic of \(\varepsilon(z)\) in [\[fb\]](#fb){reference-type="eqref" reference="fb"}, which is justified if the wavelength is close to the first Bragg resonance. To solve equation ([\[fc\]](#fc){reference-type="ref" reference="fc"}) let us represent \(E(z)\) as a sum of two waves with slowly varying amplitudes \[\label{fd} E(z)=\sqrt{\frac{q(0)}{q(z)}}\left(b_{+}(z)e^{i\int\limits^{z}_{0}q(z')dz'}+b_{-}(z)e^{-i\int\limits^{z}_{0}q(z')dz'}\right).\] The coupled wave equations for the amplitudes \(b_{\pm}(z)\) are obtained by substituting ([\[fd\]](#fd){reference-type="ref" reference="fd"}) into ([\[fc\]](#fc){reference-type="ref" reference="fc"}) and omitting the second derivatives of \(b_{\pm}(z)\) as well as the terms with double phases (for more details see ), which lead to the following system \[\label{fe} \left\{ \begin{array}{rcl} b_{+}'&=&i\frac{k^{2}}{2q(z)}\left(\mu-\frac{q^{2}(z)}{k^{2}}\right)b_{+}+i\frac{k^{2}B}{q(z)}b_{-},\\ b_{-}'&=&-i\frac{k^{2}B}{q(z)}b_{+}-i\frac{k^{2}}{2q(z)}\left(\mu-\frac{q^{2}(z)}{k^{2}}\right)b_{-}. \end{array} \right.\] The multilayer reflectivity \(R\), transmittivity \(T\) and the boundary conditions for \(b_{\pm}(z)\) are \[\begin{aligned}
R&=b_{-}(0),\quad b_{+}(0)=1,\label{fe1}\\ b_{-}(\zeta)&=0,\quad T=b_{+}(\zeta),\quad \zeta\to+\infty.\label{fe2} \end{aligned}\] The energy conservation law for the system [\[fe\]](#fe){reference-type="eqref" reference="fe"} is expressed by the relation \[\label{fe3} |b_{+}(z)|^{2}-|b_{-}(z)|^{2}=|T|^{2}=t,\] which follows directly from it and the boundary conditions [\[fe1\]](#fe1){reference-type="eqref" reference="fe1"}--[\[fe2\]](#fe2){reference-type="eqref" reference="fe2"} assuming that there is no absorption in the multilayer materials. It can be shown that the amplitude \(b=b_{-}(z)\) satisfies a second order differential equation, which follows from ([\[fe\]](#fe){reference-type="ref" reference="fe"}) \[\label{fh} b''+\frac{q'}{q}b'+\left[-iq'-\frac{k^{2}}{2}\mu+\frac{k^{4}}{4}\frac{(\mu^{2}-4B^{2})}{q^{2}}+\frac{q^{2}}{4}\right]b=0,\] where \(b=b_{-}(z)\) and the boundary conditions are \[\label{fh1} b(+\infty)=0,\quad b'(0)+i\frac{{k}^{2}}{2{q}}\left(\mu-\frac{{\rm q}^{2}}{{\rm k}^{2}}\right) b(0)=-i\frac{{k}^{2}{B}}{{q}}.\] Finally, the reflectivity \(R\) can be expressed in terms of \(b(z)\) as \[\label{fj} R=i\frac{k^{2}}{q(0)}B\frac{1}{\displaystyle i\frac{k^{2}}{q(0)}\left[\frac{q^{2}(0)}{k^{2}}-\mu\right]-\frac{b'(0)}{b(0)}}.\]
# Exact models of graded multilayers
[\[exact_form\]]{#exact_form label="exact_form"} While for an arbitrary \(q(z)\) the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} cannot be solved analytically, for a number of specific functions \(q(z)\) the analytical solution is possible. In this section three such cases will be discussed: \(q(z)=q_0(pz+\rho)\), \(q(z)=q_0(pz+\rho)^{1/3}\) and \(q(z)=q_0\tanh(pz+\rho)\), where \(p\), \(q_0\) and \(\rho\) are some parameters.
## Linear case
In the case when \(q(z)=q_0(pz+\rho)\) one can always assume that \(\rho=1\). Then introducing a new variable \[\label{ch3ex1} y=-\frac{iq_0}{2p}(pz+1)^2\] and representing the amplitude \(b(z)\) as \[\label{ch3ex2} b(z)=Cy^\nu e^{-y/2}F(y),\] where \(C\) is an arbitrary coefficient, it is possible to transform the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} into \[\label{ch3ex3} y\frac{d^2F}{dy^2}+[(2\nu+1)-y]\frac{dF}{dy}-\left[\nu+\frac{ik^2\mu}{4pq_0}\right]F=0,\] where \[\nu=\frac{ik^2}{4pq_0}\sqrt{4\mu^2-B^2}.\label{ch3ex4b}\] Equation [\[ch3ex3\]](#ch3ex3){reference-type="eqref" reference="ch3ex3"} is the confluent hypergeometric equation, which in general is known in the following form \[\label{ch3ex5} y\frac{d^2F}{dy^2}+(c-y)\frac{dF}{dy}-a F=0,\] where \(a\) and \(c\) some complex coefficients. One of solutions of [\[ch3ex5\]](#ch3ex5){reference-type="eqref" reference="ch3ex5"} is the confluent hypergeometric function of the first type or Kumer function \(M(a,c,y)={}_1F_1(a,c,y)\). However, it can be shown that the boundary condition [\[fh1\]](#fh1){reference-type="eqref" reference="fh1"} will be satisfied for a solution in the form [\[ch3ex2\]](#ch3ex2){reference-type="eqref" reference="ch3ex2"} only if the following solution of [\[ch3ex5\]](#ch3ex5){reference-type="eqref" reference="ch3ex5"} is used \[\label{ch3ex5a} U_1(a,c,y)=-\frac{\pi}{\sin\pi c}\left[\frac{M(a,c,y)}{\Gamma(c)\Gamma(1-a)}-(-y)^{1-c}\frac{M(1+a-c,2-c,y)}{\Gamma(2-c)\Gamma(c-a)}\right].\] In the end the required solution of [\[fh\]](#fh){reference-type="eqref" reference="fh"} can be written as \[\label{ch3ex6} b(z)=Cy^\nu e^{-y/2}U_1(a,c,y),\] where \[\begin{aligned}
a&=\frac{ik^2}{4pq_0}(\sqrt{\mu^2-4B^2}+\mu),\label{ch3ex7a}\\ c&=2\nu+1=\frac{ik^2}{2pq_0}\sqrt{\mu^2-4B^2}+1.\label{ch3ex7b} \end{aligned}\] Now the reflectivity \(R\) of the multilayer structure with a linearly growing inverse period \(q\) can be obtained using the formula [\[fj\]](#fj){reference-type="eqref" reference="fj"}. Substituting the solution [\[ch3ex6\]](#ch3ex6){reference-type="eqref" reference="ch3ex6"} into it leads to the following expression \[\label{ch3ex8} R=i\frac{k^{2}}{2pq_0}B\frac{1}{\displaystyle a-y_0\frac{U_1(a+1,c+1,y_0)}{U_1(a,c,y_0)}},\] where \(y_0=-iq_0/2p\). Equation [\[ch3ex8\]](#ch3ex8){reference-type="eqref" reference="ch3ex8"} can be simplified with use of the known properties of the confluent hypergeometric functions and reduced to the following formula \[\label{ch3ex9} R=i\frac{k^{2}B}{2pq_0}\frac{U_1(a,c,y_0)}{U_1(a+1,c,y_0)},\] which can then be used for practical calculations of the reflectivity.
## Cubic root case
In the case when \(q(z)=q_0(pz+\rho)^{1/3}\), similar to the linear case it can be assumed without limiting the generality that \(\rho=1\). Then the solution of [\[fh\]](#fh){reference-type="eqref" reference="fh"} can be similarly expressed through the confluent hypergeometric function. To do this let us introduce the following new variables \[\begin{aligned}
y&=\sqrt{\frac{3q_0}{2p}}\left((pz+1)^{2/3}-\frac{k^2}{q_0^2}\mu\right),\label{ch3ex10a}\\ b(z)&=Ce^{iy^2/4}U(y).\label{ch3ex10b} \end{aligned}\] Substituting [\[ch3ex10a\]](#ch3ex10a){reference-type="eqref" reference="ch3ex10a"}--[\[ch3ex10b\]](#ch3ex10b){reference-type="eqref" reference="ch3ex10b"} into the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} reduces it to the so called the parabolic cylinder (Weber) equation \[\label{ch3ex11} \frac{d^2U}{dy^2}+(y^2/4-a)U=0,\] where \[\label{ch3ex10c} a=\frac{i}{2}+\frac{3}{2}\frac{k^4}{pq_0^3}B^2.\] Solutions of the equation [\[ch3ex11\]](#ch3ex11){reference-type="eqref" reference="ch3ex11"} are Weber functions, which can be expressed through the confluent hypergeometric function. Similar to the linear case, it can be shown that only the function \(U_1\) defined in [\[ch3ex5a\]](#ch3ex5a){reference-type="eqref" reference="ch3ex5a"} has the correct asymptotic when \(z\to\infty\). Thus the required solution of the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} and the respective reflectivity can be written as \[\begin{aligned}
b(z)&=Ce^{iy^2/4}U_1(\alpha,1/2,-iy^2/2),\label{ch3ex12a}\\ R&=i\frac{k^{2}B}{q_0^2}\frac{1}{\displaystyle 1-\frac{k^2}{q_0^2}\mu}\frac{U_1(\alpha,1/2,-iy_0^2/2)}{U_1(\alpha+1,3/2,-iy_0^2/2)},\label{ch3ex12b} \end{aligned}\] where \[\begin{aligned}
\alpha&=\frac{ia}{2}+\frac{1}{4}=\frac{3i}{4}\frac{k^4}{pq_0^3}B^2,\label{ch3ex13a}\\ y_0&=\sqrt{\frac{3q_0}{2p}}\left(1-\frac{k^2}{q_0^2}\mu\right).\label{ch3ex13b} \end{aligned}\] Formula [\[ch3ex12b\]](#ch3ex12b){reference-type="eqref" reference="ch3ex12b"} is in many respects similar to the formula [\[ch3ex9\]](#ch3ex9){reference-type="eqref" reference="ch3ex9"} for the linear case. It is of interest to quantify the behavior of the obtained solution when \(p\to0\). This can be done using the WKB (quasiclassical) asymptotic of the equation [\[ch3ex5\]](#ch3ex5){reference-type="eqref" reference="ch3ex5"} (see the next section). Their application in [\[ch3ex12b\]](#ch3ex12b){reference-type="eqref" reference="ch3ex12b"} leads to the following approximate expression for the reflectivity \[\label{ch3ex13с} R\approx\frac{R_0+iP}{1+iR_0 P},\] where \(R_0\) is the reflectivity of the periodical multilayer structure with \(q=q_0\) as defined in, while the coefficient \(P\) is \[\label{ch3ex13d} P=R_0^{-i\frac{3q_0}{2p}}.\] Since it is known that inside the resonant reflectivity band \(R_0=\exp(-i\phi)\), where the phase \(\phi\) has a positive value, the coefficient \(P\) will go to zero when \(p\to0\), and \(R\to R_0\). In other words, the coupled wave approximation provides for the transition from a graded multilayer to a periodical reflective structure. For the linear case discussed in the previous subsection an analogous transition is possible.
## Hyperbolic tangent case
Finally let us consider the case when the inverse period function is \(q(z)=q_0\tanh(pz+\rho)\). The main difference with the prior two models is the fact that the function \(q\) assumes only the values from a limited interval \([q_0\tanh\rho,q_0]\). To solve the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} let us introduce the following variables \[\begin{aligned}
y&=\tanh^2(pz+\rho),\label{ch3ex14a}\\ b(z)&=Cy^{-i\zeta}(1-y)^{ \end{aligned}\] where \[\begin{aligned}
\zeta&=\frac{k^2}{4pq_0}\sqrt{\mu^2-4B^2},\label{ch3ex15a}\\ \end{aligned}\] Substituting the formulas [\[ch3ex14a\]](#ch3ex14a){reference-type="eqref" reference="ch3ex14a"}--[\[ch3ex14b\]](#ch3ex14b){reference-type="eqref" reference="ch3ex14b"} into the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} the Gauss's hypergeometric equation is obtained \[\label{ch3ex16} y(1-y)\frac{d^2U}{dy^2}+\left(c-(a+b+1)y\right)\frac{dU}{dy}-abU=0,\] where \[\begin{aligned}
a&=1-\frac{1}{4p}\left(iq_0+\frac{ik^2}{q_0}\sqrt{\mu^2-4B^2}-\frac{k^2}{q_0}\sqrt{4B^2-\left(\mu-\frac{q_0^2}{k^2}\right)^2}\right),\label{ch3ex17a}\\ b&=\frac{1}{4p}\left(iq_0-\frac{ik^2}{q_0}\sqrt{\mu^2-4B^2}+\frac{k^2}{q_0}\sqrt{4B^2-\left(\mu-\frac{q_0^2}{k^2}\right)^2}\right),\label{ch3ex17b}\\ c&=1-\frac{ik^2}{2pq_0}\sqrt{\mu^2-4B^2}.\label{ch3ex17c} \end{aligned}\] It can be shown that in order to obtain the correct asymptotic [\[fh1\]](#fh1){reference-type="eqref" reference="fh1"} when \(z\to\infty\) the solution of [\[ch3ex16\]](#ch3ex16){reference-type="eqref" reference="ch3ex16"} should be chosen in the form \(U(y)={}_2F_1(a,b,1+a+b-c,1-y)\), where \({}_2F_1\) is the Gauss's hypergeometric function. This results in the following expressions for the field amplitude and reflectivity \[\begin{aligned}
b(z)&=Cy^{-i\zeta}(1-y)^{ R&=i\frac{k^{2}B}{q_0}\frac{1}{\displaystyle i\frac{k^2}{2q_0}\left(\frac{q_0^2}{k^2}\tanh^2\rho-\mu\right)+Q},\label{ch3ex18b} \end{aligned}\] where \[\begin{gathered}
\label{ch3ex19} Q=\left(i\frac{k^2}{q_0}\sqrt{\mu^2-4B^2}(1-\tanh^2\rho)+p(c'-1)\tanh^2\rho\right)+\\ 2p\frac{ab}{c'}\tanh^2\rho(1-\tanh^2\rho)\frac{{}_2F_1(a+1,b+1,c'+1,1-\tanh^2\rho)}{{}_2F_1(a,b,c',1-\tanh^2\rho)}, \end{gathered}\] and \[\label{ch3ex20} c'=1+a+b-c=1+i\frac{k^2}{2pq_0}\sqrt{\left(\mu-\frac{q_0^2}{k^2}\right)^2-4B^2}.\] It is useful to consider the limit of [\[ch3ex18b\]](#ch3ex18b){reference-type="eqref" reference="ch3ex18b"} when \(p\to\infty\). Taking into account [\[ch3ex17a\]](#ch3ex17a){reference-type="eqref" reference="ch3ex17a"}--[\[ch3ex17c\]](#ch3ex17c){reference-type="eqref" reference="ch3ex17c"} and the Taylor expansion of \({}_2F_1\) it can be shown that in this case \({}_2F_1(a+1,b+1,2+a+b-c,1-\tanh^2\rho)\to1/\tanh^2\rho\) and \({}_2F_1(a,b,1+a+b-c,1-\tanh^2\rho)\to1\). Then the formula [\[ch3ex18b\]](#ch3ex18b){reference-type="eqref" reference="ch3ex18b"} for reflectivity turns into the know expression for the reflectivity of the periodic multilayer structure with the period \(d=\pi/q_0\) (see ). The dependence of the inverse period in the form of the hyperbolic tangent from this subsection will be used in the numerical experiments of the section [4](#sect_WKB){reference-type="ref" reference="sect_WKB"}, which are aimed at verifying the suitability of the coupled wave equations [\[fe\]](#fe){reference-type="eqref" reference="fe"} for the description of the graded multilayers.
# WKB asymptotic for the coupled wave equations {#sect_WKB}
Equation ([\[fh\]](#fh){reference-type="ref" reference="fh"}) is equivalent to the following equation \[\label{fk} \ddot{{y}}+\left[\frac{f_{0}(x)}{{p}^{2}}+\frac{f_{1}(x)}{{p}}+f_{2}(x)\right]{y}=0,\] which is the standard form for application of the WKB (Wentzel, Kramers, Brilloin) asymptotic method. In the equation [\[fk\]](#fk){reference-type="eqref" reference="fk"} \[\begin{array}{rcl} f_{0}&=&\frac{{q}^{2}}{4}+\frac{{k}^{4}}{4}\frac{(\mu^{2}-4{B}^{2})}{{q}^{2}}-\frac{{k}^{2}}{2}\mu,\\ f_{1}&=&-i\dot{{q}},\\ f_{2}&=&\frac{1}{4}\left(\frac{\dot{{q}}}{{q}}\right)^{2}-\frac{1}{2}\frac{\ddot{{q}}}{{q}}, \end{array}\] where \(z=px\), \(y=\sqrt{q}b\) and \(p\) is a formal parameter of the asymptotic expansion. The general solution of ([\[fk\]](#fk){reference-type="ref" reference="fk"}) can be expressed in the form of series by the parameter \(1/p\) as \[\label{fl} y(z)=\exp{\left(\frac{1}{p}S_{0}(x)+S_{1}+pS_{2}(x)+\ldots\right)},\] where formulas for \(S_{n}\) can be obtained by substituting [\[fl\]](#fl){reference-type="eqref" reference="fl"} into ([\[fk\]](#fk){reference-type="ref" reference="fk"}). By keeping only the first two terms in ([\[fl\]](#fl){reference-type="ref" reference="fl"}) we can obtain an approximate solution of the coupled wave equations in the following form \[\label{fm} b(z)=\frac{1}{\sqrt{q\gamma}}\left\{C_{1}\left[i(q^{2}+2\zeta)+2q\gamma\right]^{\frac{1}{2}}e^{\int\limits^{z}\gamma(z')dz'}+C_{2}\left[i(q^{2}+2\zeta)+2q\gamma\right]^{-\frac{1}{2}}e^{-\int\limits^{z}\gamma(z')dz'}\right\},\] where \[\zeta=-\frac{{k}^{2}}{2}\mu,\quad\gamma=\sqrt{-f_{0}},\quad\gamma'=\sqrt{f_{0}}.\] The solution ([\[fm\]](#fm){reference-type="ref" reference="fm"}) is only valid at large distances from the so-called turning points of equation [\[fk\]](#fk){reference-type="eqref" reference="fk"}, which coincide with the zeros \(z_{i}\) of the function \(f_{0}(z)\) at \[\label{fn} q=k\sqrt{\mu\pm2B}.\] For the sake of simplicity let us at first neglect the absorption and restrict the consideration to the real values of \(z\). It can be shown that, if \(q(z)\) changes monotonically, the equation [\[fn\]](#fn){reference-type="eqref" reference="fn"} has at most two solutions \(z_{1}\) and \(z_{2}\) lying on the positive real semi-axis. Moreover, from the point of view of the width of the reflectivity band the most important case is when both zeros are located inside the multilayer structure. This corresponds to the wavelength \(\lambda\) being within the resonant part of the reflectivity band \[\frac{2\pi\sqrt{\mu}}{q_1}<\lambda<\frac{2\pi\sqrt{\mu}}{q_2}, \label{fo}\] where it is assumed that \(q_1>q_2\). The full WKB solution of ([\[fk\]](#fk){reference-type="ref" reference="fk"}) can be constructed by extending the general solution ([\[fm\]](#fm){reference-type="ref" reference="fm"}) from one area in the complex plane of \(z\) to all other areas by using its expansions in the vicinity of the turning points \(z_{1}\), \(z_{2}\). Taking into account the boundary conditions [\[fh1\]](#fh1){reference-type="eqref" reference="fh1"} we obtain that when \(z>z_{1}\) \[\label{fp} b(z)=\frac{C}{\sqrt{{q}\gamma'}}\frac{e^{-i\int\limits_{z_{1}}^{z}\gamma'(z')dz'}}{\left[{A}+\sqrt{{A}^{2}-1}\right]^{1/2}},\] where \[A(z)=\frac{1}{2{B}}\left(\mu-\frac{{q}^{2}(z)}{k^{2}}\right).\] The expression for the amplitude \(b(z)\) in the area where \(z_{2}<z<z_{1}\) is \[\label{fq} b({z})=\frac{C}{\sqrt{{q}\gamma}}\left\{\frac{M}{\left[{A}-i\sqrt{1-{A}^{2}}\right]^{1/2}}e^{\psi_{1}}+i\left(\frac{1}{M}-M\right)\left[{A}-i\sqrt{1-{A}^{2}}\right]^{1/2}e^{-\psi_{1}}\right\},\] where \[\psi_{1}=\int\limits_{z_{2}}^{z}\gamma(z')dz',\quad M=e^{-\int\limits_{z_{2}}^{z_{1}}\gamma(z')dz'}.\] Finally, the amplitude \(b(z)\) in the area where \(z<z_{2}\) is \[\label{fr} b({z})=\frac{C}{\sqrt{{q}\gamma'}}\left\{\frac{\left[{A}-\sqrt{{A}^{2}-1}\right]^{1/2}}{M}e^{-i\psi}+i\left(\frac{1}{M}-M\right)\frac{1}{\left[{A}-\sqrt{{A}^{2}-1}\right]^{1/2}}e^{i\psi}\right\},\] where \[\psi=-\int\limits_{z_{2}}^{z}\gamma'(z')dz'.\] By substituting the solutions ([\[fp\]](#fp){reference-type="ref" reference="fp"})--([\[fr\]](#fr){reference-type="ref" reference="fr"}) into the general formula ([\[fj\]](#fj){reference-type="ref" reference="fj"}) we obtain the following expression for reflectivity inside the resonant reflectivity band ([\[fo\]](#fo){reference-type="ref" reference="fo"}) \[\label{fs} R=\frac{\displaystyle 1+i\omega\frac{e^{2i\psi}}{{A}-\sqrt{{A}^{2}-1}}}{\displaystyle {A}+\sqrt{{A}^{2}-1}+i\omega e^{2i\psi}},\] where \[\begin{aligned}
\psi&=&\int\limits_{0}^{z_{2}}\gamma'(z')dz',\nonumber\\ \omega&=&1-e^{-2\alpha},\label{fs1}\\ \alpha&=&\int\limits_{z_{2}}^{z_{1}}\gamma(z')dz',\nonumber\\ A&=&\frac{1}{2{B}}\left(\mu-{q}^{2}(0)/{k}^{2}\right). \nonumber \end{aligned}\] Formula ([\[fs\]](#fs){reference-type="ref" reference="fs"}) gives the reflectivity of a graded multilayer in terms of the material's optical constants and multilayer geometry. It can be further simplified if one is interested only in the module of the reflectivity. In the resonant reflectivity band ([\[fo\]](#fo){reference-type="ref" reference="fo"}) band and when \(|A|\gg1\) it can be shown that \[\begin{aligned}
R\approx&\frac{1}{2{A}}+i\omega e^{2i\psi},\label{ft1}\\ r=&|R|^{2}\approx-\operatorname{Im}(\frac{1}{{A}}\omega e^{2i\psi})+|\omega|^{2}e^{-4\psi''},\label{ft2} \end{aligned}\] where \[\psi''=\operatorname{Im}(\psi).\] In [\[ft2\]](#ft2){reference-type="eqref" reference="ft2"} the first term describes the oscillations of the reflectivity and the second term corresponds to the average reflectivity. Taking into account that the reflectivity band of any graded multilayer is considerably wider than the reflectivity band of a comparable periodical structure and is determined primarily by the gradient of the function \(q(z)\) rather than by the absorption, one can use the following estimates for the parameters \(\alpha\) and \(\psi''\) \[\begin{aligned}
\alpha&=&\pm\pi k^2\frac{{B}^{2}}{\operatorname{Re}(\mu)}\left|\left(\frac{dq}{dz}(\tilde{z})\right)^{-1}\right|,\label{fu}\\ \psi''&=&\beta\frac{k}{2\sqrt{\operatorname{Re}(\mu)}}\tilde{z},\label{fv} \end{aligned}\] where \(\tilde z=q^{-1}(z)\) and \(\mu=\sin^2\theta-\delta+i\beta\). This is equivalent of approximating the equation ([\[fn\]](#fn){reference-type="ref" reference="fn"}) for the zeros of \(f_0(z)\) as \(q(z)=k(\sqrt{\mu}\pm B/\sqrt{\mu})\). Let us note that in ([\[fu\]](#fu){reference-type="ref" reference="fu"}) and ([\[fv\]](#fv){reference-type="ref" reference="fv"}) it is also assumed that the absorption is small, i.e. that \(\beta\ll\delta\ll 1\). The formula [\[ft2\]](#ft2){reference-type="eqref" reference="ft2"} can be further simplified by averaging over the oscillations and by substituting \(B\) for \(|B|\). The latter can be done because the absorption is small. Then the parameter \(\omega\) becomes a real number and for the averaged reflectivity we will have \[r\approx\omega^{2}e^{-4\psi''}. \label{fv1}\] Finally, to access the accuracy of the approximate formulas obtained in this section we conducted a number of numerical experiments. We chose the model function \(q(z)=q_0\tanh(pz+\rho)\), where \(q_0\), \(p\) and \(\rho\) are some parameters, and for which the equation [\[fh\]](#fh){reference-type="eqref" reference="fh"} has the exact solution as it was demonstrated above. Then we compared the reflectivity calculated using the formula [\[fs\]](#fs){reference-type="eqref" reference="fs"} with that obtained by the direct solution of the Helmholtz equation ([\[fa1\]](#fa1){reference-type="ref" reference="fa1"}) in a multilayer coating using the matrix method for the dielectric susceptibility dependence defined in [\[fa5\]](#fa5){reference-type="eqref" reference="fa5"}. To apply the matrix method one has to express the bilayer thicknesses \(d_n\) in terms of function \(q(z)\). This can be done by using the definition of function \(q(z)\) to obtain the positions of the bilayer interfaces \(z_n\) \[2\int\limits_{0}^{z_{n}}q(z)dz=2\pi n, \label{fz3}\] which are extremums of \(\cos(2\int\limits_{0}^{z_{n}}q(z)dz)\) in ([\[fb\]](#fb){reference-type="ref" reference="fb"}). After that the bilayer thicknesses can be calculated as \(d_n=z_n-z_{n-1}\), where \(n=1,2,\ldots\). All calculations with the analytical formulas as well as by the direct matrix method were performed using the specially written FORTRAN 95 program codes. The X-ray optical constants were taken from the Henke tables. Two such comparisons for \(Cr/C\) multilayer structures in the hard and soft ("carbon window") X-ray spectral ranges are shown in Fig.[\[f1\]](#f1){reference-type="ref" reference="f1"}--[\[f2\]](#f2){reference-type="ref" reference="f2"}, which also depict the reflectivities calculated using the analytical formula [\[ch3ex18b\]](#ch3ex18b){reference-type="eqref" reference="ch3ex18b"}. As it is seen from Fig.[\[f1\]](#f1){reference-type="ref" reference="f1"}--[\[f2\]](#f2){reference-type="ref" reference="f2"} the approximate formula ([\[fs\]](#fs){reference-type="ref" reference="fs"}) is in a reasonably good agreement with the direct solution of the Helmholtz equation ([\[fa1\]](#fa1){reference-type="ref" reference="fa1"}) sometimes reproducing even small features of rather complicated profiles of \(r(\lambda)\), although it diverges and is clearly inapplicable near the boundaries of the resonant reflectivity band. On the other hand, the formula [\[ch3ex18b\]](#ch3ex18b){reference-type="eqref" reference="ch3ex18b"} gives the results, which are even in a better agreement with the direct solution. The accuracy of the analytical formula increases when the gradient of the inverse period decreases (\(p\to0\)) or when the reflectivity band is narrow (as in Fig.[\[f2\]](#f2){reference-type="ref" reference="f2"}). This is exactly what is expected from the general theory formulated in the sections [\[sect_coupled\]](#sect_coupled){reference-type="ref" reference="sect_coupled"} and [4](#sect_WKB){reference-type="ref" reference="sect_WKB"}.
# Inverse problem
[\[sect_inv\]]{#sect_inv label="sect_inv"} It is very useful to compare results of the perturbation theory, which is also known as the kinematic approximation (see for details), and results given by the WKB asymptotics. Let us first observe that the formula ([\[fv1\]](#fv1){reference-type="ref" reference="fv1"}) in case \(|B|\leqslant1\) behaves asymptotically as \(r\approx4\alpha^{2}e^{-4\psi''}\sim B^{4}\). On the other hand, when \(|B|\leqslant1\) the kinematic approximations is also applicable, from which it follows that the reflectivity should be proportional to \(B^{2}\), but not to \(B^{4}\), i.e. to the square of modulation of the dielectric permittivity in the multilayer structure (see ). To rectify this problem we will use the following formula for the parameter \(\omega\) instead of the expression [\[fs1\]](#fs1){reference-type="eqref" reference="fs1"} \[\omega\approx\sqrt{\tanh(\alpha)}. \label{fz4}\] In the limit when \(\alpha\to\infty\) formula [\[fz4\]](#fz4){reference-type="eqref" reference="fz4"} coincides with [\[fs1\]](#fs1){reference-type="eqref" reference="fs1"} simultaneously making the reflectivity [\[fv1\]](#fv1){reference-type="eqref" reference="fv1"} behave correctly in the kinematic approximation, i.e. when \(\alpha\to0\). The numerical experiments have also demonstrated that use of the formula ([\[fv1\]](#fv1){reference-type="ref" reference="fv1"}) may lead to the values of the reflectivity that are somewhat elevated over those obtained numerically by the matrix method. Since this can cause errors in the inverse problem solution the following modified formula for the reflectivity is used here \[r\approx\frac{r_{old}}{1+r_{old}}, \label{fz5}\] where \[r_{old}=\tanh(\alpha)e^{-4\psi''},\] which alleviates this problem. Now the formula ([\[fz5\]](#fz5){reference-type="ref" reference="fz5"}) together with the Bragg condition \(q(\tilde{z})=\sqrt{\mu}k\) can be used to derive a differential equation for the function \(q(z)\), which solves the inverse problem of constructing a multilayer coating with a given dependence of the reflectivity on the wavenumber \(k\) or grazing angle \(\theta\). For the wavelength dependence it has the following form \[\frac{dz}{dq}=\mp\frac{(\operatorname{Re}(\mu))^{2}}{2\pi |B|^{2}}\frac{1}{q^{2}}\ln\left(\frac{1-\frac{r}{1-r}e^{4\psi''}}{1+\frac{r}{1-r}e^{4\psi''}}\right) \label{fz2}\] with the initial condition \(z(q_{0})=0\) and where \[r=r\left(\frac{q}{\sqrt{\operatorname{Re}(\mu)}}\right)\] and \(r(x)\) is a pre-specified function of \(k\). For the grazing angle dependence the differential equation is \[\frac{dz}{dq}=\mp\frac{q^2}{2\pi k^{4}}\frac{1}{B^2} \ln\left(\frac{1-\tilde r(q)(1+e^{4\psi''})}{1-\tilde r(q)(1-e^{4\psi''})}\right), \label{fz6}\] with the same initial condition \(z(q_{0})=0\) and where \[\tilde r(q)=r\left(\frac{q^2}{k^2}+1-\operatorname{Re}(\mu_0)\right)\] and \(r(x)\) is a given function of \(\sin\theta\). The plus sign in [\[fz2\]](#fz2){reference-type="eqref" reference="fz2"} and [\[fz6\]](#fz6){reference-type="eqref" reference="fz6"} should be chosen to obtain the increasing function \(q(z)\), and the minus sign--the decreasing function \(q(z)\). In other words, every inverse problem has at least two solutions. To demonstrate the ability of the developed method to solve real world inverse problems a number of numerical experiments in different wavelength ranges and with diverse shapes of the reflectivity curve has been carried out. Their results are shown in Fig.[\[f3\]](#f3){reference-type="ref" reference="f3"}--[\[f4\]](#f4){reference-type="ref" reference="f4"}. The differential equation [\[fz2\]](#fz2){reference-type="eqref" reference="fz2"} for the growing \(q(z)\) was solved numerically by an explicit Euler method using a specially written FORTRAN 95 program code. The obtained functions \(q(z)\) were converted to the layer thicknesses using the equation [\[fz3\]](#fz3){reference-type="eqref" reference="fz3"}, which then were used to calculate the multilayer reflectivities (also shown in Fig.[\[f3\]](#f3){reference-type="ref" reference="f3"}--[\[f4\]](#f4){reference-type="ref" reference="f4"}) by the direct matrix method as it was done in the numerical experiments of the previous section. The X-ray optical constants were again taken from the Henke tables. It is clear from Fig.[\[f3\]](#f3){reference-type="ref" reference="f3"}--[\[f4\]](#f4){reference-type="ref" reference="f4"} that the method proposed here is capable of constructing blueprints of the graded multilayers, whose reflectivities approximate the target reflectivities reasonably well except for the observed quasiperiodic oscillations. The agreement is appreciably better in the hard X-ray region than in the soft "carbon window" spectral range. The calculations also showed (see Table [1](#t1){reference-type="ref" reference="t1"}) that the graded multilayers might have the integrated (over wavelength) reflectivity at least several times larger than that of a periodical structure. Thus the developed theory can be used as a basis for the solution of the inverse problem and maximization of the integral reflectivity.
# Maximization of the integral
The reflectivity in ([\[fv1\]](#fv1){reference-type="ref" reference="fv1"}) depends on the function \(q^{-1}(q)\), which is the inverse of the function \(q(z)\), and on its first derivative. Therefore we can identify the reflectivity \(r\) with the Lagrange function known from the classical mechanics, the integral reflectivity--with the action, \(q^{-1}(q)\)--with a generalized coordinate and the wavelength or grazing angle with the time. The maximization problem for the integral reflectivity in an interval of wavelengths (or grazing angles) is then reduced to the Lagrange equation \[\frac{\partial r}{\partial z}=\frac{d}{dq}\frac{\partial r}{\partial z'_{q}} \label{ch3eq2h}\] with the boundary conditions \[z(q_1)=0,\quad z(q_2)=L,\label{ch3eq2ha}\] where \(q_1\) and \(q_2\) are the boundary values of the function \(q(z)\) and \(L\) is the multilayer's thickness. From the equation ([\[ch3eq2h\]](#ch3eq2h){reference-type="ref" reference="ch3eq2h"}) the following second order differential equation is obtained \[\begin{gathered}
\label{ch3eq2i} \frac{d^2z}{dq^2}=\frac{1}{2a^3q^3}\frac{4c\frac{\tanh \frac{2}{aq^2}\frac{1-r_0}{\tanh \end{gathered}\] where the plus or minus sign corresponds to an increasing or decreasing function \(q(z)\), respectively. In the equation ([\[ch3eq2i\]](#ch3eq2i){reference-type="ref" reference="ch3eq2i"}) \[\begin{aligned}
r_0&=&\tanh a&=&\pi\frac{|B|^2}{(\mu')^2},\quad c=\frac{\gamma}{2}\frac{1}{(\mu')^2}.\nonumber \end{aligned}\] The derivatives of \(a\) and \(c\) assume non-zero values only if the material dispersion is present. For the maximization of the integral over an angular interval an equation similar to [\[ch3eq2i\]](#ch3eq2i){reference-type="eqref" reference="ch3eq2i"} can be obtained. Equation [\[ch3eq2i\]](#ch3eq2i){reference-type="eqref" reference="ch3eq2i"} is quite complicated and in the general case can be solved only numerically. It is, however, possible to estimate the upper limit of the achievable integral reflectivity in rather a simple way. Let us assume that \(R(\lambda)\) (or, respectively, \(R(\theta)\)) has a table-like profile in the wavelength interval \([\lambda_{min},\lambda_{max}]\) (or in the grazing angle interval \([\theta_{min},\theta_{max}]\)) \[r(\lambda)=r_c(\theta(\lambda-\lambda_{min})-\theta(\lambda-\lambda_{max})), \label{ch3eq2g3}\] where \(r_c\) is a positive number less than 1. For [\[ch3eq2g3\]](#ch3eq2g3){reference-type="eqref" reference="ch3eq2g3"} the inverse problem can be solved as described in the section [\[sect_inv\]](#sect_inv){reference-type="ref" reference="sect_inv"}. The maximum or minimum value of \(q\) will be determined by the location of zero in the argument of logarithm in [\[fz2\]](#fz2){reference-type="eqref" reference="fz2"} or [\[fz6\]](#fz6){reference-type="eqref" reference="fz6"}, respectively. After the maximal inverse period range \(\Delta q_{max}=|q_{max}-q_{0}|\) is found, one can find using the Bragg condition the maximal reflectivity bandwidth in the wavelength or angle domain for the specified value of \(r_c\), and then the maximal integral reflectivity. The maximum inverse period range \(\Delta q_{max}\) can be estimated by a direct integration in the equations ([\[fz2\]](#fz2){reference-type="ref" reference="fz2"}) or ([\[fz6\]](#fz6){reference-type="ref" reference="fz6"}). Neglecting for simplicity the material dispersion we have \[\Delta q_{max}\leqslant-\pi \bar{q}\frac{|B|^{2}}{\gamma}\frac{1}{\mu'}\int\limits_{s}^{1}\frac{du}{u}\frac{1}{\ln\left[\frac{1-u}{1+u}\right]}, \label{ch3eq2g4}\] where \(\bar{q}\) is some average value of \(q\) and the lower integration boundary is defined as \(s=r_{c}/(1-r_{c})\). Using the formula ([\[ch3eq2g4\]](#ch3eq2g4){reference-type="ref" reference="ch3eq2g4"}) as well as the Bragg condition ([\[fn\]](#fn){reference-type="ref" reference="fn"}) (taken at \(B=0\)), both wavelength and angle bandwidths can be expressed as \[\begin{aligned}
\Delta\lambda&\leqslant&-\pi\bar{\lambda}\frac{|B|^{2}}{\gamma}\frac{1}{\mu'}\int\limits_{s}^{1}\frac{du}{u}\frac{1}{\ln\left[\frac{1-u}{1+u}\right]},\label{ch3eq2g6}\\ \Delta\theta&\leqslant&-\pi\bar{\theta}\frac{|B|^{2}}{\gamma}\frac{1}{\mu'}\int\limits_{s}^{1}\frac{du}{u}\frac{1}{\ln\left[\frac{1-u}{1+u}\right]},\label{ch3eq2g7} \end{aligned}\] respectively, where parameters \(B\), \(\mu\) and \(\gamma\) should be taken in some middle point \(\bar{\lambda}\) or \(\bar{\theta}\). Let us consider the case when \(r_c\ll1\) and, therefore, \(s\ll1\). Then the common integral in ([\[ch3eq2g6\]](#ch3eq2g6){reference-type="ref" reference="ch3eq2g6"}) and ([\[ch3eq2g7\]](#ch3eq2g7){reference-type="ref" reference="ch3eq2g7"}) can be evaluated as \[I=\int\limits_{s}^{1}\frac{du}{u}\frac{1}{\ln\left[\frac{1-u}{1+u}\right]}\approx\frac{1}{2}\left[\frac{1}{s}-1-\frac{1}{\sqrt{3}}\left(\frac{\pi}{6}-\mbox{atan}\frac{s}{\sqrt{3}}\right) \right], \label{ch3eq2g8}\] which diverges when \(s\to0\). This is not surprising as the bandwidth is expected to become progressively wider when the reflectivity decreases. If \(r_c\) were zero, the bandwidth would be infinite. Let us multiply ([\[ch3eq2g8\]](#ch3eq2g8){reference-type="ref" reference="ch3eq2g8"}) by \(r_c\) and then substitute the result into [\[ch3eq2g6\]](#ch3eq2g6){reference-type="eqref" reference="ch3eq2g6"} or [\[ch3eq2g7\]](#ch3eq2g7){reference-type="eqref" reference="ch3eq2g7"}. The outcome is an estimate of the maximal integral reflectivity over a wavelength or grazing angle range \[\begin{aligned}
r_{int,\lambda}&\leqslant\frac{\pi}{2}\bar{\lambda}\frac{|B|^{2}}{\gamma\mu'}\left[1-r_c\left(2+\frac{\pi}{2\sqrt{3}}\right)+\frac{r_c}{\sqrt{3}}\operatorname{atan}\frac{r_c}{\sqrt{3}(1-r_c)}\right],\label{ch3eq2g9}\\ r_{int,\theta}&\leqslant\frac{\pi}{2}\bar{\theta}\frac{|B|^{2}}{\gamma\mu'}\left[1-r_c\left(2+\frac{\pi}{2\sqrt{3}}\right)+\frac{r_c}{\sqrt{3}}\operatorname{atan}\frac{r_c}{\sqrt{3}(1-r_c)}\right].\label{ch3eq2g9a} \end{aligned}\] The average wavelength or grazing angle is assumed to be equal to \(\bar{\lambda}=\sqrt{\lambda_1\lambda_2}\) or \(\bar{\theta}=\sqrt{\theta_1\theta_2}\), respectively, where \(\lambda_1\) and \(\lambda_2\) or \(\theta_1\) and \(\theta_2\) are the wavelengths or grazing angles, corresponding to the boundaries of the respective reflectivity bands. From the formulas ([\[ch3eq2g9\]](#ch3eq2g9){reference-type="ref" reference="ch3eq2g9"}) and ([\[ch3eq2g9a\]](#ch3eq2g9a){reference-type="ref" reference="ch3eq2g9a"}) it follows that, when \(r_c\) decreases, the integral reflectivity \(r_{int}\) increases reaching the maximum at \(r_c=0\). Thus the limiting maximal value of the integral reflectivity can be obtained by substituting \(r_c=0\) into [\[ch3eq2g9\]](#ch3eq2g9){reference-type="eqref" reference="ch3eq2g9"} or [\[ch3eq2g9a\]](#ch3eq2g9a){reference-type="eqref" reference="ch3eq2g9a"} \[\begin{aligned}
r_{int,\lambda}^{lim}&\leqslant\frac{\pi}{2}\bar{\lambda}\frac{|B|^{2}}{\gamma\mu'},\label{ch3eq2g10}\\ r_{int,\theta}^{lim}&\leqslant\frac{\pi}{2}\bar{\theta}\frac{|B|^{2}}{\gamma\mu'}.\label{ch3eq2g10a} \end{aligned}\] The maximal values of the integral reflectivities calculated using ([\[ch3eq2g10\]](#ch3eq2g10){reference-type="ref" reference="ch3eq2g10"}) are shown in Table [1](#t1){reference-type="ref" reference="t1"} in the last column, while the respective average values of the wavelength or grazing angle in the previous column. From Table [1](#t1){reference-type="ref" reference="t1"} it can be seen that the integral reflectivities of real structures are 10--40% smaller than the estimated maximal values, especially in the soft X-ray range.
Formulas ([\[ch3eq2g10\]](#ch3eq2g10){reference-type="ref" reference="ch3eq2g10"}) and ([\[ch3eq2g10a\]](#ch3eq2g10a){reference-type="ref" reference="ch3eq2g10a"}) can be also used to obtain the optimal value of \(\beta\), at which the maximal integral reflectivity is reached. Noting that ([\[ch3eq2g10\]](#ch3eq2g10){reference-type="ref" reference="ch3eq2g10"}) and ([\[ch3eq2g10a\]](#ch3eq2g10a){reference-type="ref" reference="ch3eq2g10a"}) depend on \(\beta\) only through a factor \[m=|B|^{2}/\gamma=\frac{|\varepsilon_{1}-\varepsilon_{2}|^{2}}{4\pi^{2}}\frac{\sin^{2}\pi\beta}{\operatorname{Im}(\varepsilon_{2})+\beta(\varepsilon_{1}-\varepsilon_{2})}.\] Differentiating the latter expression with respect of \(\beta\) and equaling the result to zero we can obtain the following equation for \(\beta_{max}\) \[\tan\pi\beta_{max}=2\pi(\beta_{max}+g), \label{ch3eq2g11}\] where \(g=\operatorname{Im}\varepsilon_{2}/\operatorname{Im}(\varepsilon_{1}-\varepsilon_{2})\). When one material in a multilayer structure has the absorption much large than the second material and therefore \(g\ll1\), the solution of [\[ch3eq2g11\]](#ch3eq2g11){reference-type="eqref" reference="ch3eq2g11"} is \(\beta_{max}\approx0.37\). Let us note that while ([\[ch3eq2g11\]](#ch3eq2g11){reference-type="ref" reference="ch3eq2g11"}) was obtained for the reflectivities having table-like shapes, it will be approximately valid for any shape resembling the table-like profile.
# Conclusion
An analytical theory of the graded multilayer coatings based on the coupled wave approximation was developed. A number of exact solutions of the coupled wave equations for different dependencies of the inverse period on the depth were found. A series of numerical experiments demonstrated a good agreement between the graded multilayer reflectivities calculated using the obtained exact solutions with the calculations by the exact matrix method. The coupled wave amplitudes were then approximated by WKB asymptotics and approximate formulas for the multilayer reflectivity were obtained. The numerical experiments then showed that the use of the WKB asymptotics leads to the reflectivities that are in a satisfactory agreement with the matrix method calculations especially within the resonant band. It was also revealed that a significant widening of the reflectivity band as compared to the periodical multilayers is achievable. As the next step differential equations for solution of the inverse problem for graded multilayers in the wavelength or angle domain were derived based on the developed analytical theory. A number of numerical experiments, which demonstrated the suitability of this approximate method for the real world inverse problem solution in different wavelength ranges, was conducted. They demonstrated a good agreement of the obtained inversions with the target reflectivities. The general problem of the integral reflectivity maximization was then discussed and a general second order differential equation for its solution was derived from the variational principle. Finally, a formula, which estimates the upper bound of the achievable integral reflectivity, was obtained and compared to the results of the numerical experiments. To summarize, the developed analytical theory of graded multilayers can be used to design non-periodic X-ray multilayers in a wide range of wavelengths (from 0.1 to 10 nm) and grazing angles. The inversions obtained with it can be further rectified using standard numerical optimization methods, for instance, the fastest descent method. This means that the presented theory can potentially facilitate the development of the broadband X-ray reflective optics for such applications as the "water window" and "carbon window" microscopes as well as for the X-ray telescopes working in the hard part of the X-ray spectrum. | {'timestamp': '2021-09-14T02:15:18', 'yymm': '2109', 'arxiv_id': '2109.05330', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05330'} |
# Equivalence between definitions of precategories {#sec:precat-equivalent-definitions}
We prove the equivalence between the equational and the enriched definition of precategories: | {'timestamp': '2021-09-14T02:17:02', 'yymm': '2109', 'arxiv_id': '2109.05369', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05369'} |
# Introduction
Consider a commodity whose market exhibits three types of agents: buyers, sellers, and middlemen. Each seller owns one indivisible unit; and each buyer seeks to purchase one unit (from any of the sellers) in exchange for money. Units may not be homogeneous, *i.e.*, a buyer may have different valuations for the respective units owned by two distinct sellers. We assume that utility is transferable between all agents; and this allows the use of cooperative games with transferable utility (or TU games, for short). A given buyer and a given seller may trade directly, or they may use the services of a middleman. For example, in the real estate market, a seller may or may not use a realtor facilitating the sale of her house. In financial markets, brokers provide their service to investors (in exchange for a fee); and each investor may or may not hire a broker. As is common in these applications, we assume that a middleman may serve multiple buyer-seller pairs. Markets with middlemen have been studied in different contexts (search and matching models, general equilibrium model, etc.). was the first work to study the activity of middlemen in search markets. In agents can search for matches on their own, or they can resort to a middleman who mediates between agents of opposite sides to facilitate their pairing. investigates competition between middlemen when direct trade between buyers and sellers is available. He showed that direct trade has a negative effect on the market power of middlemen. consider a market with search frictions and a monopolistic middleman where buyers and sellers bargain over the surplus. study a model in which sellers and buyers have heterogeneous tastes. They showed that middlemen are better off if they have a multi-unit inventory of differentiated products. In their seminal paper, used a TU game to model a two-sided housing market where there are \(m\) buyers and \(n\) sellers. In their setting, each buyer is interested in buying at most one house and each seller has one house for sale. Each buyer has \(n\) valuations (one for each house) each seller has a reservation value for her house. The valuation matrix represents the joint surplus generated by each pair formed by a buyer and a seller. For this market situation the associated TU game, the so-called *assignment game*, is defined. They studied a solution concept, the *core*, which is the set of allocations that cannot be improved upon by any coalition. They showed that the core of an assignment game is always non-empty and has a lattice structure. Moreover, and proved that there exists a core allocation at which each buyer attains his/her marginal contribution to the grand coalition (the buyer-optimal core allocation) and there exists a core allocation at which each seller attains his/her marginal contribution to the grand coalition (the seller-optimal core allocation). Multi-sided matching markets may in general have an empty core under transferable utility ([@kw82]). Thus, the remarkable results obtained for two-sided markets cannot be generalized to all multi-sided markets. Several authors have examined conditions (on the structure of the market) allowing to show the non-emptiness of the core. introduced a subclass of multi-sided matching markets where valuations are obtained from a supermodular function. She proved that any game in this subclass has a non-empty core. Some other authors have shown that matching markets exhibiting some additivity property have a non-empty core (see for instance [@q91]; [@t13]; [@aetal16]). Among different multi-sided matching market models, there is a growing literature on matching markets with middlemen. introduces a three-sided matching game with middlemen, the so-called *supplier-firm-buyer game*. In this model, a buyer and a seller (supplier) can trade only through a middleman (firm). Hence, unlike our model, a mixed-pair of buyer-seller cannot generate any surplus without a middleman. The author showed that the class of supplier-firm-buyer game is balanced. consider a model of three-sided matching markets in which middlemen can mediate at most one trade between a buyer and seller. Buyers and sellers are allowed to trade directly as well as trade through a middleman. Unlike our model, middlemen incur a matching cost and moreover, the associated TU game only considers the matching situations with triplets of buyer-middleman-seller. Moreover, the present paper also relaxes the condition on middlemen by allowing each of them to serve multiple buyer-seller pairs. In a recent paper, study a hybrid model of two-sided and multi-sided matching markets. They consider a two-sided model with buyers and sellers that are not disjoint. There exists a so-called central player who can act both as a buyer and as a seller. In their model, the central player has to be present for a trade between a buyer-seller pair. Otherwise, a trade cannot be realized. Hence, the central player has veto power and, as explained by the authors, their model thus induces a veto game ([@b16]). The present work takes a game-theoretical approach to matching markets with middlemen. We consider a class of three-sided matching market in which buyers and sellers can trade directly or indirectly through middlemen. Each seller owns an object to sell and each buyer wants to acquire at most one object. A trade between a mixed-pair of a buyer and a seller can be mediated by at most one middleman, meanwhile any given middleman can mediate trades between multiple buyer-seller pairs. Utility is transferable and quasi-linear in money. We assume that a direct trade between a buyer-seller pair is more costly (and therefore generates a lower surplus) than when a middleman is involved (thanks to her knowledge of the market, the middleman allows the buyer and the seller to lower their search costs and travel costs). Given a buyer-seller pair, the surplus generated by their exchange varies depending on the middleman serving them. As mentioned before, every buyer (seller) can trade with at most one seller (buyer), whereas any given middleman may serve multiple buyer-seller pairs (note that any such pair is served by at most one middleman). In order to study the core and its structure, we propose a simple procedure allowing to compute the worth of the grand coalition in any matching market with middlemen. Precisely, we construct an associated two-sided assignment market where the valuation of every buyer-seller pair is obtained by taking the maximum surplus that they can achieve either by a direct trade between themselves or by an indirect trade brokered by any of the middlemen in the market. In a similar fashion, introduce the maximum assignment game for a given collection of assignment games where any given coalition attains the maximum possible value among the given collection of games. However, in their case, the two authors observed that the maximum assignment game need not be an assignment game, and it may not even be superadditive. Our main results are described as follows. First, we show that an optimal matching for a matching market with middlemen can always be constructed from an optimal matching of the associated two-sided market, and vice versa. Moreover, the maximum total surplus in the two markets are equal (Proposition [\[prop:opt_matchings\]](#prop:opt_matchings){reference-type="ref" reference="prop:opt_matchings"}). Second, we prove that the core of a market with middlemen is always non-empty by showing that the set of payoff vectors composed of a core allocation for the two-sided assignment market and zero payoffs to all middlemen is precisely the subset of the core of the market with middlemen where all middlemen payoffs are zero (Theorem [\[TheoremCoreNonempty\]](#TheoremCoreNonempty){reference-type="ref" reference="TheoremCoreNonempty"}). Furthermore, we prove that there exists a buyer-optimal allocation, that is, a core allocation that each and every buyer (weakly) prefers to all other core allocations. Likewise, there exists a seller-optimal core allocation (Theorem [\[TheoremBuyOptimal\]](#TheoremBuyOptimal){reference-type="ref" reference="TheoremBuyOptimal"}). Moreover, as in the standard two-sided model, our results guarantee that, at the buyer-optimal (seller-optimal) core allocations, each buyer (seller) achieves her marginal contribution to the grand coalition. Interestingly, we provide an example showing that, in general, there exists no middleman-optimal core allocation (Example [\[ex:no_mc_middlemen\]](#ex:no_mc_middlemen){reference-type="ref" reference="ex:no_mc_middlemen"}): all middlemen do not necessarily achieve their maximum core payoffs simultaneously. Finally, we characterize the core in terms of competitive equilibrium payoffs (Theorem [\[thm:Core_euqivalence_CE\]](#thm:Core_euqivalence_CE){reference-type="ref" reference="thm:Core_euqivalence_CE"}). The paper is organized as follows. In Section [2](#sec:prel){reference-type="ref" reference="sec:prel"} some preliminaries are given. Section [3](#sec:mode){reference-type="ref" reference="sec:mode"} introduces the model and explores the structure of its outcomes, the matchings. In Section [4](#sec:core){reference-type="ref" reference="sec:core"} we prove the non-emptiness of the core and prove that there exists a side-optimal core allocation for the buyer side and also for the seller side in the market. In contrast, by means of an example we demonstrate that there need not exist a middleman-optimal core allocation. In Section [5](#sec:CE){reference-type="ref" reference="sec:CE"} we establish the coincidence between the core and the set of competitive equilibrium payoff vectors. Section [6](#sec:conc){reference-type="ref" reference="sec:conc"} concludes.
# Preliminaries {#sec:prel}
A *cooperative game with transferable utility* (or TU game) is a pair \((N,v)\) where \(N\) is a non-empty, finite set of *players (or agents)* and \(v:2^{N}\rightarrow\mathbb{R}\) is a *coalitional function* satisfying \(v(\emptyset)=0\). The number \(v(S)\) is the *worth* of the coalition \(S\subseteq N\). Whenever no confusion may arise as to the set of players, we will identify a TU game \((N,v)\) with its coalitional function \(v\). Given a game \(v\), a *payoff allocation* (or allocation) is a tuple \(x\in\mathbb{R}^{N}\) representing the players' respective allotments. The total payoff of a coalition \(S\subseteq N\) is denoted by \(x(S)=\sum_{t\in S} x_{t}\) if \(S\neq \emptyset\) and \(x(\emptyset)=0\). In a game \(v\), an allocation \(x\) is called *efficient* if \(x(N)=v(N)\), *individually rational* if \(x_{t}=x(\{t\})\geq v(\{t\})\) for all \(t\in N\), and *coalitionally rational* if \(x(S)\geq v(S)\) for all \(S\subseteq N\). The *core* of \(v\), denoted by \(Core(v)\), is the set of coalitionally rational and efficient payoff allocations. A game is called *balanced* if it has a non-empty core, and *totally balanced* if all the subgames, i.e. the game restricted to the non-empty coalitions, are balanced. A totally balanced game \(v\) is balanced and also *superadditive*, i.e. \(v(S\cup T)\geq v(S)+v(T)\) for any coalitions \(S,T\subseteq N\) such that \(S\cap T=\emptyset\). We call *marginal contribution* of a player \(t\in N\) in the game \(v\) the quantity \(mc_t(v)=v(N)-v(N\setminus \{t\})\). It is well known that the marginal contribution is an upper bound of the payoffs attainable in the core for a player, i.e. \(x_t\leq mc_t(v)\) for all \(x\in Core(v)\) and \(t\in N\), but this bound is not necessarily sharp.
# Matching markets with middlemen {#sec:mode}
We consider a three-sided market where there are three disjoint sets of agents: the set of buyers \(B=\{b_1,b_2,\ldots, b_I\}\), the set of middlemen \(M=\{m_1, m_2, \ldots, m_J\}\), and the set of sellers \(S=\{s_1,s_2,\ldots, s_K\}\). Note that the cardinalities \(I,J,K\) of these respective sets may differ. We call \(B\) (or \(S\)) the *short side of the market* if it holds that \(I\leq K\) (or \(|S|\leq |B|\)). Let \(N=B\cup M\cup S\) be the set containing all agents. In this market, each buyer-seller pair \((i,k)\in B\times S\) can trade directly with each other, or indirectly through some middleman \(j\in M\) which results in a trade involving the triple \((i,j,k)\in B\times M\times S\). Each seller owns one unit of good and each buyer seeks to buy at most one unit of good. Although a trade between each buyer-seller pair can be mediated by at most one middleman, any given middleman can mediate trades between multiple buyer-seller pairs. That is to say, each \(j\in M\) can potentially serve the entire market by brokering as many trades as the cardinality of the short side of the market. A market with middlemen can thus be described by specifying two non-negative matrices: (a) a two-dimensional matrix \(A=(a_{ik})_{\substack{i\in B \\ k\in S}}\) giving the joint monetary *surplus generated by every mixed pair* \((i,k)\in B\times S\) if they trade directly, and (b) a **three-dimensional** non-negative matrix \(\hat{A}=(\hat{a}_{ijk})_{\substack{i\in B\\ j\in M \\k\in S}}\) representing the *joint surplus generated by a trade between buyer \(i\) and seller \(k\) that is mediated by middleman \(j\)*. It will often be convenient to represent the three-dimensional surplus matrix \(\hat{A}\) as an array of its two-dimensional layer submatrices indexed by the middlemen. Formally, \(\hat{A}=[A^{(j)}\in \mathbb{R}^{B\times S}: j\in M]\) with elements \(a^{(j)}_{ik}=\hat{a}_{ijk}\) for all \((i,j,k)\in B\times M\times S\). For sake of unified notation, we denote the two-dimensional surplus matrix \(A\) as \(A^{(0)}\) with elements \(a^{(0)}_{ik}=a_{ik}\) for all \((i,k)\in B\times S\). We will assume that \[\label{Assmatrices} a^{(j)}_{ik}=\hat{a}_{ijk}\geq a_{ik}=a^{(0)}_{ik}, ~\forall (i,j,k)\in B\times M\times S,\] that is to say, a direct trade between a buyer-seller pair entails higher search costs than when a middleman is involved. Hence, the total surplus of a trade with a middleman is greater than or equal to that of a trade without a middleman. A *market with middlemen* is fully described by a tuple of the type \(\gamma=(B,M,S,A,\hat{A})\). Since the sets \(B,M,S\) are given and fixed, we will often describe such a market with middlemen by simply specifying a pair of matrices \((A,\hat{A})\) satisfying ([\[Assmatrices\]](#Assmatrices){reference-type="ref" reference="Assmatrices"}). Call *basic coalition* any subset of \(N\) that is either a singleton \(\{i\}\), or a pair \(\{i,k\}\) such that \(i\in B\) and \(k\in S\), or a triple \(\{i,j, k\}\) such that \(i\in B\), \(j\in M\) and \(k\in S\). Moreover, let \(\mathcal{B}^{N}=\{\{i,j,k\}\mid i\in B, j\in M, k\in S\}\cup\{\{i,k\}\mid i\in B, k\in S\}\cup \{\{i\}\mid i\in N\}\) be the collection of all basic coalitions. Furthermore, for all \(T\subseteq N\), denote by \(\mathcal{B}^{T}\) the set of basic coalitions that have all their agents in \(T\), that is, \(\mathcal{B}^{T}=\{E\in\mathcal{B}^{N} \mid E\subseteq T\}\). Denote by \(B_T\), \(M_T\), and \(S_T\) the set of buyers, middlemen, and sellers in coalition \(T\), respectively.
Remark that conditions (i)-(iii) in Definition [\[defmatching\]](#defmatching){reference-type="ref" reference="defmatching"} say that a buyer (seller) must belong to exactly one basic coalition in the collection \(\mu\). It is possible for a middleman to belong to multiple basic triples of \(\mu\) (since she may mediate multiple trades). However, as stated in (iv), a middleman appearing in a singleton of \(\mu\) should not belong to any other element of \(\mu\). With a slight abuse of notation, we write \(k= \mu(i)\) and \(i= \mu(k)\) for all \((i,k)\in B\times S\) such that \(\left[\{i,k\}\in \mu \mbox{ or } \{i,j,k\}\in \mu \mbox{ for some } j\in M\right]\). We also write \(\mu(t)=t\) for all \(t\in N\) such that \(\{t\}\in \mu\). Let \(\mathcal{A}(T)\) denote the set of \(T\)-matchings. Observe that a \(T\)-matching \(\mu\) induces disjoint groups of buyer-seller pairs that trade via the same middleman. With a slight abuse of notation, we shorthand the subsets containing only one agent from each side of the market as an array in which the ordered specifies the type of the agents: \((i,k)\) means \(\{i,k\}\) with \(i\in B\) and \(k\in S\); similarly, \((i,j,k)\) means \(\{i,j,k\}\) with \(i\in B\), \(j\in M\), and \(k\in S\). We call the buyers in the set \(B^{\mu}_{j}=\{i\in B_T: (i,j,k)\in\mu \textrm{ for some } k\in S_T\}\) and the sellers in the set \(S^{\mu}_{j}=\{k\in S_T: (i,j,k)\in\mu \textrm{ for some } i\in B_T\}\) the *partners* of middleman \(j\in M_T\) in \(T\)-matching \(\mu\). Let \(M^{\mu}_{+}\) denote the set of those middlemen in \(T\) who are involved in some trading triplet under \(\mu\). Denote by \(B^{\mu}_{0}\) (\(S^{\mu}_{0}\)) the set of those buyers (sellers) in coalition \(T\), who are not partners of any middleman but are involved in some direct trade under \(\mu\), as if they were partners of a fictitious middleman denoted by 0. Finally, denote the set of buyers, middlemen, and sellers in \(T\) who are singletons in \(\mu\) by \(B^{\mu}_{s}\), \(M^{\mu}_{s}\), and \(S^{\mu}_{s}\) respectively. Obviously, \(M^{\mu}_{s}\) together with the singletons \(\{j\}\) \((j\in M^{\mu}_{+}=M_T\setminus M^{\mu}_{s})\) form a partition of \(M_T\), \(B^{\mu}_{s}\) together with the partner sets \(B^{\mu}_{j}\) \((j\in M^{\mu}_{+}\cup \{0\})\) form a partition of \(B_T\), and \(S^{\mu}_{s}\) together with the partner sets \(S^{\mu}_{j}\) \((j\in M^{\mu}_{+}\cup \{0\})\) form a partition of \(S_T\). Moreover, the union of these three partitions form a partition of coalition \(T\), called *the \(\mu\)-induced partition* of \(T\). Notice that \(\mu\) induces a (complete) matching, denoted by \(\mu^{(j)}\), between the partner sets \(B^{\mu}_{j}\) and \(S^{\mu}_{j}\) of each (real or fictitious) middleman \(j\in M^{\mu}_{+}\cup \{0\}\). Consequently, \(|B^{\mu}_{j}|=|S^{\mu}_{j}|\) for any \(j\in M^{\mu}_{+}\cup \{0\}\). Note that a market \(\gamma=(A,\hat{A})\) induces a TU game \(v_\gamma\) where the worth of every coalition \(T\) is given by \[\label{EqworthT} v_\gamma(T)=\max_{\mu\in \mathcal{A}(T)} \left[ \sum_{(i,k)\in \mu}a_{ik}+\sum_{(i,j,k)\in \mu} \hat{a}_{ijk} = \sum_{j\in M^{\mu}_{+}\cup \{0\}} \sum_{(i,k)\in \mu^{(j)}} a^{(j)}_{ik} \right]\] Note from ([\[EqworthT\]](#EqworthT){reference-type="ref" reference="EqworthT"}) that all coalitions \(T\) consisting of players of the same side (including singleton coalitions) are worthless, that is, \(v_\gamma(T)=0\). A matching \(\mu\in\mathcal{A}(T)\) will be called \(T\)-*optimal* in the market \(\gamma\) if \(v_\gamma(T)=\sum\limits_{(i,k)\in \mu}a_{ik}+\sum\limits_{(i,j,k)\in \mu} \hat{a}_{ijk}\), that is, if \(\mu\) solves the problem stated in ([\[EqworthT\]](#EqworthT){reference-type="ref" reference="EqworthT"}). Since \(\mathcal{A}(T)\) is non-empty and finite, remark that there always exists (at least) one \(T\)-optimal matching in \(\gamma\). Given any \(T\in 2^N\), we denote by \(\mathcal{A}^*_\gamma(T)\) the set of \(T\)-optimal matchings in the market \(\gamma\). We call *optimal matching* any \(N\)-optimal matching in \(\gamma\). The following example illustrates the notions developed in this section.
In the next section we examine the core of the TU game \(v_\gamma\) associated with the matching market with middlemen \(\gamma\). We will show in particular that this game, called *middlemen market game*, \(v_\gamma\) is always totally balanced.
# The core of a market with middlemen {#sec:core}
Given any market with middlemen \(\gamma=(B,M,S, A, \hat{A})\), one can define the matrix \(A^*=(a^*_{ik})_{\substack{i\in B \\ k\in S}}\) by \[\label{EqAstar} a^*_{ik}=\max_{j\in M} \hat{a}_{ijk}, ~\forall (i,k)\in B\times S.\] Note from ([\[Assmatrices\]](#Assmatrices){reference-type="ref" reference="Assmatrices"}) and ([\[EqAstar\]](#EqAstar){reference-type="ref" reference="EqAstar"}) that \(a^*_{ik}\) gives the highest surplus possible in a trade involving buyer \(i\) and seller \(k\). Moreover, for all \((i,k)\in B\times S\), we will use the notation \(m(i,k)\) to refer to the *lowest-label middleman* \(j\in M\) such that \(a^*_{ik}=\hat{a}_{ijk}\), that is to say, \(m(i,k)=\min \argmax\limits_{j\in M} \hat{a}_{ijk}=\min\{j\in M: \hat{a}_{ijk}=\max_{h\in M} \hat{a}_{ihk}\}\). Note that due to our simplifying assumption ([\[Assmatrices\]](#Assmatrices){reference-type="ref" reference="Assmatrices"}), \(m(i,k)\in M\) is well defined for any buyer-seller pair \((i,k)\in B\times S\). Thus, for any market \(\gamma=(A, \hat{A})\), one can define the standard (two-sided) assignment market \(\gamma^*=(B,S,A^*)\), where \(A^*\) is given by ([\[EqAstar\]](#EqAstar){reference-type="ref" reference="EqAstar"}). Note that a matching \(\nu\) in \(\gamma^*\) is a partition of \(B\cup S\) into singletons and mixed pairs \(\{i,k\}\) such that \(i\in B, k\in S\). We write \(\nu(t)=t\) for all \(t\in B\cup S\) such that \(\{t\}\in \nu\). In addition, we write \(\nu(i)=k\) and \(\nu(k)=i\) for all \((i,k)\in B\times S\) such that \(\{i,k\}\in \nu\). A matching \(\nu\) in \(\gamma^*\) is optimal if \(\sum\limits_{\{i,k\}\in \nu}a^*_{ik}\geq \sum\limits_{\{i,k\}\in \nu'}a^*_{ik}\), for all matchings \(\nu'\) in \(\gamma^*\). We connect the matchings, in particular the optimal matchings, of the respective markets \(\gamma\) and \(\gamma^*\). For expositional simplicity, we only consider the grand coalition, the concepts are analogously defined and the statements are straightforwardly derived for any subcoalition. Let \(\mu\) be a matching in a market with middlemen \(\gamma=(B,M,S,A,\hat{A})\). As it induces partitions of the set of buyers \(B\) and sellers \(S\), and matchings \(\mu^{(j)}\) (\(j\in M\cup\{0\}\)) between the partner sets for each middleman which are pairwise disjoint for different middlemen, the union \(\bigcup_{j\in M\cup\{0\}} \mu^{(j)}\) augmented with the singletons in \(B^{\mu}_{s}\) and \(S^{\mu}_{s}\) defines a matching between \(B\) and \(S\). We denote it by \(\mu^*\). The value of \(\mu\) in the market \(\gamma\) is clearly less than or equal to the value of \(\mu^*\) in the two-sided market \(\gamma^*\), that is, \[\label{EqMuMuStar} \mu_\gamma(B\cup M\cup S)=\sum\limits_{(i,k)\in \mu} a_{ik} + \sum\limits_{(i,j,k)\in \mu} \hat{a}_{ijk} \leq \sum\limits_{(i,k)\in \mu^*} a^*_{ik} = \mu^*_{\gamma^*}(B\cup S).\] Conversely, if \(\sigma\) is a matching for the two-sided market \(\gamma^*\), then \(\sigma^{\triangle}=\{(i,m(i,k), k): (i,k)\in \sigma\} \cup \left\{ \{t\} \in \sigma \right\} \cup \left\{ \{j\}:j\in M \mbox{ s.t. } j\neq m(i,k), \forall (i,k) \in \sigma\right \}\) is a matching for the market with middlemen \(\gamma\). The value of \(\sigma^{\triangle}\) in the market \(\gamma\) is clearly the same as the value of \(\sigma\) in the two-sided market \(\gamma^*\), that is \[\label{EqSigmaStarSigma} \sigma_{\gamma^*}(B\cup S) = \sum\limits_{(i,k)\in \sigma} a^*_{ik} = \sum\limits_{(i,m(i,k),k)\in \sigma^{\triangle}} \hat{a}_{im(i,k)k} = \sigma^{\triangle}_\gamma(B\cup M\cup S).\] Based on ([\[EqMuMuStar\]](#EqMuMuStar){reference-type="ref" reference="EqMuMuStar"}) and ([\[EqSigmaStarSigma\]](#EqSigmaStarSigma){reference-type="ref" reference="EqSigmaStarSigma"}), we derive the following relations between the optimal matchings and the optimum total surplusses in the two markets.
Proposition [\[prop:opt_matchings\]](#prop:opt_matchings){reference-type="ref" reference="prop:opt_matchings"} shows that one can always construct an optimal matching in the market with middleman \(\gamma\) by first finding an optimal matching of the associated two-sided market \(\gamma^*\). Next, we reconsider Example [\[Ex:Illustritaive\]](#Ex:Illustritaive){reference-type="ref" reference="Ex:Illustritaive"} to illustrate Proposition [\[prop:opt_matchings\]](#prop:opt_matchings){reference-type="ref" reference="prop:opt_matchings"}.
The following result proves that the TU game associated with a market with middlemen is always totally balanced.
Any two-sided assignment market exhibits two distinguished core allocations, namely the buyer-optimal allocation and the seller-optimal allocation. Under the buyer-optimal (seller-optimal) allocation all buyers (sellers) simultaneously achieve their maximum core payoff and all sellers (buyers) simultaneously achieve their minimum core payoff. showed that, in the two-sided assignment market, the existence of these optimal allocations is a result of the "lattice structure" of the core. Moreover, and showed that the buyers (sellers) achieve their marginal contribution to the grand coalition at the buyer-optimal (seller-optimal) allocation. Nevertheless, this property does not extend to the arbitrary multi-sided markets. study a special case of multi-sided markets where each of the \(m\) sides has an optimal core allocation under which all agents of that side achieve their marginal contribution. Our next result states that, in a matching market with middlemen, there exists an optimal core allocation for buyers (sellers): under their optimal allocation, all buyers (sellers) simultaneously achieve their marginal contribution.
Remark that Theorem [\[TheoremBuyOptimal\]](#TheoremBuyOptimal){reference-type="ref" reference="TheoremBuyOptimal"} claims the existence of a buyer-optimal (seller-optimal) allocation, but not that of a middleman optimal allocation. Indeed, it is not true in general that there exists an allocation where all middlemen achieve their highest payoff in the core. The following example illustrates this point.
# Core and competitive equilibria {#sec:CE}
The aim of this section is to study the relationship between core and competitive equilibria in matching markets with middlemen. defines competitive equilibrium prices and proves their existence for any assignment problem (see also [@ss71]). extends the coincidence between core and competitive equilibria for the classical three-sided assignment markets where buyers are forced to acquire exactly one item of each type. In a similar fashion, generalizes the equivalence result for the generalized three-sided assignment markets where buyers can buy at most one good of each type. In both extensions, the existence of a competitive equilibrium is guaranteed whenever the core is non-empty. Consider any market with middlemen where the set of buyers is \(B=\{b_1, \ldots, b_I\}\), the set of middlemen is \(M=\{m_1, \ldots, m_J\}\), and the set of sellers is \(S=\{s_1,\ldots,s_K\}\). Assume that buyers and sellers trade through the competitive market with the presence of middlemen and agents in the market are price-takers. Each buyer \(i\in B\) demands at most one unit of the good, each seller \(k\in S\) offers one unit for sale (recall that units owned by different sellers may be heterogeneous). Assume that buyer \(i\) values the good of seller \(k\) at \(h_{ik}\), and the production cost of the good for seller \(k\) is \(c_{k}\). If buyer \(i\) and seller \(k\) trade directly, the transaction (search) cost \(t_{ik}\), is incurred by buyer \(i\). If buyer \(i\) instead hires middleman \(j\) and ends up purchasing the object owned by seller \(k\), then the transaction cost \(t_{ijk}\) is incurred by buyer \(i\); and middleman \(j\in M\) incurs the mediation cost \(c_{j}^{ik}\). From of our assumption that the search cost of a direct trade is lower than that of a mediated trade, it comes that \(t_{ik}\geq t_{ijk}+c_{j}^{ik}\). Let \(p_{k}\) be the price demanded by seller \(k\) for her unit; and assume that middleman \(j\in M\) charges a fee \(p_{j}^{ik}\) to buyer \(i\) when the latter uses \(j\)'s services to purchase the unit owned by seller \(k\). Note that middlemen need not charge the same fee for each possible buyer-seller trade. That is, it may happen that \(p_{j}^{ik}\neq p_{j}^{i'k'}\) when \(j\) is mediating the respective pairs \((i,k)\) and \((i',k')\) (with the possibility of having either \(i=i'\) or \(k=k'\)). If the transaction between buyer \(i\) and seller \(k\) is realized through middleman \(j\), then the utility of buyer \(i\) is given by \(h_{ik}-t_{ijk}-p^{ik}_{j}-p_{k}\), the benefit of seller \(k\) is \(p_{k}-c_{k}\), and the benefit of middleman \(j\) is \(p^{ik}_{j}-c^{ik}_{j}\). Thus, the total surplus is \(h_{ik}-t_{ijk}-p^{ik}_{j}-p_{k}+p^{ik}_{j}+p_{k}-c^{ik}_{j}-c_{k}=h_{ik}-t_{ijk}-c_{j}^{ik}-c_{k}\). If \(h_{ik}-t_{ijk}-c_{j}^{ik}-c_{k}<0\), no transaction will be realized since a transaction will go through only if it gives a non-negative utility to each of the three agents \(i\), \(j\) and \(k\). Thus, for all \((i,j,k)\in B\times M\times S\), let \(\hat{a}_{ijk}=\max\{0,h_{ik}-t_{ijk}-c^{ik}_{j}-c_{k}\}\) denote the surplus generated when a transaction is realized between buyer \(i\) and seller \(k\) through middleman \(j\). Similarly, when the transaction is realized directly between buyer \(i\) and seller \(k\), the utility of buyer \(i\) is \(h_{ik}-t_{ik}-p_{k}\), the benefit of seller \(k\) is \(p_{k}-c_{k}\), and hence the total surplus is \(h_{ik}-t_{ik}-p_{k}+p_{k}-c_{k}=h_{ik}-t_{ik}-c_{k}=a_{ik}\). If \(h_{ik}-t_{ik}-c_{k}<0\), no transaction will be realized between buyer \(i\) and seller \(k\). Thus, for all \((i,k)\in B\times S\), let \(a_{ik}=\max\{0,h_{ik}-t_{ik}-c_{k}\}\) denote the surplus generated when a transaction is realized directly between buyer \(i\) and seller \(k\). Hence, this detailed market situation can be summarized by a tuple simply giving the set of buyers, the set of middlemen, the set of sellers, and the two matrices with generic terms \(a_{ik}\) and \(\hat{a}_{ijk}\) defined above. That is to say, the TU game \((N, v_{\gamma})\) associated with this market is defined precisely by the characteristic function \(v_\gamma\) given in ([\[EqworthT\]](#EqworthT){reference-type="ref" reference="EqworthT"}). We want to show that each core allocation can be obtained as the result of trading at competitive prices. To do so, we need some definitions allowing to introduce the notion of competitive price vector. A price vector \(p\in \mathbb{R}^{B\times M\times S}_{+}\times \mathbb{R}^{S}_{+}\) contains the specific, possibly differentiated prices, of the mediation services for each buyer-middleman-seller configuration as well as the undifferentiated prices of the goods. Given a matching market with middlemen \(\gamma\), a *feasible price vector* is \(p\in \mathbb{R}^{B\times M\times S}_{+}\times \mathbb{R}^{S}_{+}\) such that \(p^{ik}_{j}\ge c^{ik}_{j}\) for all \(j\in M\) and \(p_{k}\ge c_{k}\) for all \(k\in S\). The set of basic coalitions that contain buyer \(i\in B\) is \(\mathcal{B}^{i}=\{E\in\mathcal{B}^{N}\mid i\in E\}\). Let \(w^{i}(E)=h_{ik}-t_{ijk}\) be the valuation of buyer \(i\) for \(E=\{i,j,k\}\) and \(w^{i}(E)=h_{ik}-t_{ik}\) be the valuation of buyer \(i\) for \(E=\{i,k\}\). Observe the relation \[\label{valuation-basic} v_\gamma(E)=\max \left\{0,\, w^i(E)-c(E\setminus \{i\} \right\}\] for any basic coalition \(E\in\mathcal{B}^{i}\) containing buyer \(i\). Next, for each feasible price vector \(p\in \mathbb{R}^{B\times M\times S}_{+}\times\mathbb{R}^{S}_{+}\) we introduce the *demand set* of each buyer \(i\in B\).
Observe that a competitive equilibrium consists of a set of prices and an \(N\)-matching where each buyer maximizes her utility under the assignment of \(N\)-matching and prices. Moreover, middlemen and sellers are competitive, in the sense that no middleman mediates a trade unless she can charge a fee (service price) at least equal to her cost and no seller agrees to sell her good without receiving at least her cost. If a pair \((p,\mu)\) is a competitive equilibrium, then we say that the price vector \(p\) is a *competitive equilibrium price vector* and the \(N\)-matching \(\mu\) is a *compatible matching*. The corresponding payoff vector for a given pair \((p,\mu)\) is called *competitive equilibrium payoff vector*. This payoff vector is \((x(p,\mu),y(p,\mu),z(p,\mu))\in\mathbb{R}^{B}\times\mathbb{R}^{M}\times\mathbb{R}^{S}\), defined by \[\begin{aligned}
x_{i}(p,\mu)&=&w^{i}(E^{\mu(i)})-p(E^{\mu(i)}\setminus \{i\})\mbox{ where } i\in E^{\mu(i)}\in\mu \quad\mbox{ for all }i\in B,\\ y_{j}(p,\mu)&=&\sum\limits_{\{i,j,k\}\in\mathcal{\mu}} p_{j}^{ik}-\sum\limits_{\{i,j,k\}\in\mathcal{\mu}} c_{j}^{ik} = p_{j}(\mu)-c_{j}(\mu) \quad \mbox{for all } j\in M,\\ z_{k}(p,\mu)&=&p_{k}-c_{k} \quad \mbox{ for all } k\in S. \end{aligned}\] Notice the dependence of the aggregated service prices (fees) \(p_{j}(\mu)\) and the aggregated service costs \(c_{j}(\mu)\) on the matching \(\mu\). We denote the set of competitive equilibrium payoff vectors of market \(\gamma\) by \(\mathcal{CE}(\gamma)\). We now study the relationship between the core of \(\gamma=(B,M,S,A,\hat{A})\) and the set of competitive equilibrium payoff vectors. First, we show that an \(N\)-matching \(\mu\) is an optimal matching whenever it constitutes a competitive equilibrium with a feasible price vector \(p\).
Now, we can provide the main result of this section. We establish the equivalence between the core and the set of competitive equilibrium payoff vectors.
We have shown that the core and the set of competitive equilibrium payoff vectors coincide under the assumption that middlemen need not charge the same price for two different buyer-seller trade. Next example shows that if we consider the case where middlemen charge a fixed price for each buyer-seller trade that they mediate, then a core allocation need not to be supported by competitive prices.
# Concluding remarks {#sec:conc}
We have considered a class of multi-sided matching markets where a trade between buyer-seller pairs can be realized with or without middlemen. We allow a middleman to serve the entire market by mediating as many trades as the size of the short side of the market while buyer-seller pairs can also trade directly. We have associated a classical two-sided assignment market with a matching market with middlemen by taking for each buyer-seller pair the maximum surplus that this pair can achieve with the costless help of middlemen. We have shown that the non-empty core of this associated two-sided assignment market can be embedded in the core of the matching market with middlemen by allocating zero payoff to all middlemen. For these markets we have introduced an associated TU game, thereby we have extended the classical (two-sided) assignment markets of to a special multi-sided case. We have shown that every matching market with middlemen has a non-empty core. In addition, we have proved that there exists a buyer-optimal and a seller-optimal core allocation for every matching market with middlemen. Unlike in other extensions previously studied, it is shown that all buyers (sellers) achieve their marginal contribution simultaneously at the buyer-optimal (seller-optimal) core allocation. In addition, we have provided an example to show that it is not the case for middlemen: in general there does not exist an allocation that every middleman weakly prefers to any other allocation in the core. Finally, we have studied the relationship between the core and the set of competitive equilibria. We have established the coincidence between the core and the set of competitive equilibrium payoff vectors. A possible direction for further research is to study the relationship between the core and another set-wise solution concept, the bargaining set. proved the equivalence between the classical bargaining set of, a set-wise solution concept based on bargaining possibilities of players, for two-sided assignment games (see also [@s08] for related results on other partitioning games). generalized this result among others to a larger class known as (quasi)-hyperadditive games. For multi-sided matching markets, the coincidence result between the classical bargaining set and the core is exhibited only to the class of supplier-firm-buyer games ([@as18]). Nevertheless, the methods used in the aforementioned papers do not seem to carry over to our model and we leave exploring the relationship between the bargaining set and the core for future research.
[^1]: Without loss of generality, let us denote buyers, middlemen, and sellers by their indices in an order of buyer-middleman-seller. For instance, \(\{1,2,1\}\) for the coalition \(\{b_{1},m_{2},s_{1}\}\). | {'timestamp': '2022-01-25T02:47:08', 'yymm': '2109', 'arxiv_id': '2109.05456', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05456'} |
# Introduction {#sec:introduction}
In the last decade, advancement in robotic grasping has enabled robots to automatically grasp a never-before-seen range of objects. However, most of the works on grasp synthesis still assume specific object properties such as uniform friction or rigidity. These assumptions do not hold for multi-material or deformable objects and can lead to unsuccessful grasping in real-world scenarios. Grasping non-rigid objects, on the other hand, is difficult because objects deform under interaction forces meaning that the 3-D contact locations also depend on the forces exerted on the object. Furthermore, the effect of the deformation varies across deformable objects and tasks. In some scenarios, such as grasping a water bottle, it is useful to minimize the object's deformation not to dislodge the liquid. While for other objects, such as the triangular-shaped object shown in one can take advantage of the deformation to grasp them successfully. To date, most of the existing works only focus on minimizing the object deformation. Although there exist few works that take advantage of the deformation, they mainly focus on proposing control strategies given an initial grasp configuration. Thus, it is still an open question of how object stiffness affects the choice of grasp configuration and how to harness object deformation to generate better grasps. To address the aforementioned open issues, we propose to incorporate stiffness as an additional input to a state-of-the-art deep grasp planning pipeline as shown in . Our system generates a grasp candidate and corresponding grasp quality for every pixel given an input depth image and a *stiffness image*. The model outputs can be reprojected into 3D space when combined with depth information, allowing a robot to execute a generated grasp in the real world. We qualitatively evaluated the proposed grasp synthesis method on a Franka Emika Panda In summary, the main contributions of this paper are:
- The first generative stiffness-aware deep grasping approach that adapts the grasp location depending on the object's stiffness.
- The first stiffness-dependent image-based grasping dataset consisting of labeled top-down grasps on objects with varying stiffnesses.
- A thorough empirical evaluation of the proposed method presenting, both in simulation and on real hardware, improvements in terms of grasp ranking and grasp success rate over a method that ignores stiffness.
# Related work {#sec:related_work}
To put our work in context, we next review three complementary viewpoints, grasping of deformable objects, data-driven grasp synthesis, and simulation of deformable object interactions.
## Grasping Deformable Objects
Most recent research on deformable object manipulation has mainly focused on manipulating cloth items and ropes. Grasping deformable objects remains a sparsely explored research area with the majority of works focusing on formulating quality metrics to quantify the goodness of a grasp or proposing control strategies. Most approaches for grasping deformable objects aim to minimize the deformation. For a particular grasp, the minimization can be performed on-line by employing a control strategy that regulates the force at each contact. To plan grasps, minimum deformation can be achieved by placing fingers at locations with maximal local stiffness, determined e.g. using simulation. The deformation can also be integrated as an additional component of a wrench-space grasp quality metric. In contrast to minimizing deformation, some works have proposed to utilize the object deformation, similar to this paper. Analytical grasp planning approaches following this line of study include bounded force closure which guarantees force closure under a bounded external force, and deform closure which generalizes form closure to deformable objects with frictionless contact. In the on-line case, finger displacements can be regulated in order to retain force closure, originally proposed for planar objects and later extended to 3D by using [fem]{acronym-label="fem" acronym-form="singular+short"} to continuously model changes in shape and contact geometry during object lifting. Although utilized object deformations for stable grasping, they focus on regulating either force or displacement of the fingers given an initial grasp configuration. This is in contrast to our work which focuses on the choice of the grasp configuration by taking advantage of the object deformation.
## Data-driven Grasp Synthesis
Rapid advances in deep learning research recently have changed the paradigm of robotic grasping from analytical methods to data-driven ones. The main reason for this paradigm shift is that data-driven methods have been proved to be able to generate grasps that typically achieve a high grasp success rate on a wide range of objects in just a matter of seconds, much faster compared to analytical methods. For example, Mahler *et al.* used a dataset consisting of millions of synthetic antipodal top-grasp to train a Grasp Quality Convolutional Neural Network (GQ-CNN) model that computes the probability of success of grasps from depth images. The GQ-CNN was further improved through the use of on-policy data and a [fcn]{acronym-label="fcn" acronym-form="singular+short"} structure called FC-GQ-CNN. The [fcn]{acronym-label="fcn" acronym-form="singular+short"} structure has recently been found to perform well in grasp synthesis, having the ability to generate dense, pixel-wise predictions for an input image efficiently. Although the aforementioned works achieve impressive results on rigid objects, none of them explicitly investigate the usability on deformable objects especially 3D deformable objects. In this work, we tackle this problem by incorporating object deformation into a state-of-the-art deep data-driven grasping planning pipeline.
## Soft-body Simulation
Data-driven grasping requires training data, which in this work we generate in simulation. Simulating dynamics of deformable objects relies heavily on their geometric representations, for instance, particle representation is a good choice for simulating the dynamics of fluids. Yin *et al.* presents three primary deformable object modelling approaches, [mss]{acronym-label="mss" acronym-form="singular+short"}, [pbd]{acronym-label="pbd" acronym-form="singular+short"}, and [fem]{acronym-label="fem" acronym-form="singular+short"}, and their limitations. In this work, we decided to use [fem]{acronym-label="fem" acronym-form="singular+short"} because it is often used to model 3D objects such as food and tissues and, compared to other modeling approaches, offers a more physically accurate representation of a deformable object in a continuous domain at the expense of computational cost. Some robotic simulators that support [fem]{acronym-label="fem" acronym-form="singular+short"} are PyBullet, SOFA, and NVIDIA's recent version of the Isaac Gym simulator, which supports soft body simulation through the NVIDIA Flex backend. Among the aforementioned simulators, Isaac Gym is chosen as it combines the advantages of the other two. Specifically, similar to SOFA, Isaac Gym includes a co-rotational linear model for precision in modeling and simulating the object deformation under interaction. Furthermore, similar to PyBullet, the Isaac simulator also provides the capability to integrate robot-related functions, making it easier to build robotic applications. Huang *et al.* also provides a grasping framework to automatically perform and evaluate grasp tests on an arbitrary target object. We use this framework in our work to generate training data and to test grasps.
# Experiments and Results {#sec:exp_and_res}
The experiments address the following three questions:
- Can *Def-GG-CNN*
- Is *Def-GG-CNN*
- Can *Def-GG-CNN*
- How does the stiffness information affect the choice of the grasp configuration?
## Grasping in Simulation
To investigate the quality of synthesized grasps in simulation, we evaluated the approach on two sets of objects: 7 common objects shown in , and 28 adversarial objects from the recent EGAD! test dataset. For each object and stiffness, we evaluated the five grasps with the highest quality on a *shake task* shows the simulation result of both test sets. We can see that the proposed approach that takes stiffness input into account achieves a higher grasp success rate across all object sets and disturbances. It is noteworthy, however, that although we trained the network on the quality metrics from the *shake task* Focusing on the *shake task* We also evaluated the models on the multi-material object 7 shown in , where the stiffness of the red part could differ from the blue part. The result showed that if we assumed object 7 is entirely rigid, the method could not generate any good grasps on it. However, if we assumed the red part was softer than the blue part, we could generate successful grasps that usually aimed for the softer area of the object. This simple example demonstrates the benefit of planning stiffness-aware grasps on irregular-shaped multi-material objects.
## Sensitivity Analysis
To examine the robustness of the results of our approach in the presence of uncertainty, we conducted a sensitivity analysis where we introduce uncertainty to the input stiffness images by varying the stiffness parameter, *i.e.*, Young's modulus across an error range of \[-60%, +60%\]. We then evaluated the generated grasps on the Common test set under the *shake task* shows the result of the sensitivity analysis. We can see that the grasp success rate decreases consistently with increasing error, but that small errors in the range of 5-15% have only limited negative effect on the performance. This experiment indicates that our method is robust against some errors in expected stiffness, which suggests potential for real-world application even when the stiffness is not precisely known.
## Grasp Transfer to Physical Robot
To investigate how well the synthesized grasps perform in real world, we evaluated the grasp success rate on a Franka Emika Panda We used an Intel RealSense D345 camera mounted to the robot's wrist to capture the RGB-D image. In addition to depth image, we also need to provide a stiffness image of the object. To do this, we segmented the object from the scene by subtracting the background and the table from the image and then assigning the same stiffness value to each pixel that the object occupied. We manually set the magnitude of stiffness for each object according to its perceived stiffness. The best grasp pose is then computed using the proposed method. The robot executed the best grasp by moving to a pre-grasp position approximately 25 mm above the grasp. Then, the robot moves linearly downwards until reaching the grasp pose or contact with the table is detected. From there, the robot closes its gripper, lifts the object, performs a predefined trajectory, and finally places an object at the goal position. A grasp is successful if the robot can pick the object and move it without dropping it. Otherwise, it is unsuccessful. To single out the effect stiffness input has on grasp performance for each object, we ran the experiments with three different stiffnesses: the correct one, only high, only low. For each object and stiffness, we randomly placed it ten times and evaluated the best grasp candidate. In total, this setup amounts to 420 grasps on 14 objects. The experimental results are presented in . The results show that with the correct stiffness information, the grasp success rate of the proposed approach is approximately 17% higher than when the stiffness is assumed to be either high or low. This result indicates that grasps generated on rigid objects do not necessarily transfer successfully to deformable objects and vice versa. For instance, if we assumed the deformable objects, such as objects 8, 9, 11, and 14, were rigid, many generated grasps were on specific parts of them, such as the arms or legs of the toy or wheel of the car. These grasps usually picked the object successfully but then, due to the deformation, dropped it when the robot started to accelerate. The same experiment on objects 6, 7 showed that grasps generated on the top or bottom of the object usually failed due to the elasticity of the object. If we instead assumed rigid objects, such as 1, 3, 4, 5, were soft, the network generated pinch grasps that failed due to collisions with the object. Some failed grasps are shown in a. One interesting finding was that given the correct stiffness information, the method was able to generate different grasping strategies depending on the stiffness of the object, as shown in . Specifically, in the case of the soft sponge a, the proposed method learned that the grasp quality is high across the whole objects thanks to their deformation, which in turn, enables pinch grasps. While in the case of a hard sponge b, the high-quality grasp tends to be generated at the center of the object, and the grasp width is almost as big as the object in order to successfully cage the object.
## Discussion
All experimental results show the benefit of generating stiffness-aware grasps. By comparing the proposed approach to the case where the stiffness information is ignored, we see that the proposed approach achieves higher grasp success rates. The primary reason for the difference in performance is that the object stiffness facilitates learning where to generate grasps that minimize the slippage caused by the deformation. If the object's stiffness was ignored, the network generated the same grasps regardless of the object stiffness. Together, these result backs the claim made in that grasps do not transfer well between rigid and deformable objects. Therefore, incorporating object stiffness in robotic grasping pipelines is beneficial when dealing with a wide range of unknown objects. Another interesting finding is that our approach can generate different grasp types such as pinch or cage grasps depending on the stiffness, even though there was no pinch grasps in the training dataset. This behavior is shown on object 6 in . Specifically, the sponge with a low Young's modulus admits pinching behavior where the grasp press on the object and pinch, while the hard sponge only admits caging grasps. One potential reason for such behavior is that the proposed network learned that the grasp quality is almost uniform across soft objects thanks to their deformation. Similar behaviors were also reported in where data was collected from 14 robots over the course of two months. However, it is worth pointing out that our approach learned to produce the same behavior on a completely synthetic dataset with orders of magnitude fewer data. Furthermore, our proposed approach provides more meaningful insights regarding the relationship between object deformation and grasps.
# Conclusions and future work {#sec:conclusions}
Grasping deformable objects has not been as well studied as rigid object grasping due to complexity in the modeling and simulating the dynamic behavior of such objects. However, with the rapid development of physics-based simulators that support soft bodies, the research gap between rigid and deformable objects is shrinking. To leverage the capability of such simulators and to challenge the rigidity assumption that has dominated robotic grasping, we presented an approach to synthesize grasps on objects with varying stiffness by a deep neural network trained on purely synthetic data. The key idea in this work is to integrate the object stiffness property into the grasp planning pipeline to study the relationship between the object deformation and the generated grasps. To train the proposed network, we generated our own training dataset using the Isaac Gym simulator. We demonstrated the performance of the generated grasps through experiments in both simulation and real-world scenarios on a wide range of objects with varying sizes, shapes, and stiffness. The results show a clear improvement in grasp success rate when taking stiffness property into account. Furthermore, the proposed approach shows the ability to generate different grasp strategies depending on object stiffness. The generalizability to objects with non-uniform stiffness remains open although the method should be able to account for all variability captured in the training data, making the simulation quality and efficiency a central bottleneck. The idea of exploiting deformations for grasping opens many interesting avenues. Our proposed method employs FEM simulations for which the computational cost may limit their use for general-purpose systems. Investigating the possibility to devise analytical quality measures that would exploit the deformations is an important future work.
# Method {#sec:method}
## Network
Our solution is based on the GG-CNN because it is orders of magnitude smaller than other recent grasping networks, thus it is faster to train and evaluate the network, while achieving state-of-the-art results in grasping rigid objects. We propose *Def-GG-CNN*
## Grasp map representation
Each pixel in the grasp map **G** represents a 4-dof grasp. We use the same representation of **G** as defined in. As shown in , the grasp map **G** consist of three images: grasp quality **Q**, orientation \(\boldsymbol\phi\), and gripper width **W**. **Q** denotes the quality of a grasp centered at each pixel. The quality of a grasp is a scalar value between \[0,1\], where the higher the value, the better the grasp. \(\boldsymbol\phi\) is the orientation image, representing the pixel-wise orientation of a grasp around the image normal. Because an antipodal grasp is symmetric beyond 180 degrees, we limit the orientation between \[\(-\pi/2,~\pi/2\)\] radians. Finally, **W** is the width image that describes the pixel-wise gripper width from \[0, 150\] pixels. We transform the pixel-dependent gripper width to real-world units using the measured depth and the camera parameters.
# Problem Formulation
This work addresses the problem of generating antipodal grasps on unknown objects with different stiffnesses lying on a supporting surface. The goal is to calculate a grasp for each pixel in the depth image while taking into account object stiffness. More formally, we train a model \(\mathcal{M}\) that takes as input a depth image \(\mathbf{I_d}\) and a stiffness image \(\mathbf{I_s}\), and produces a grasp map **G** that incorporates grasp quality and grasp parameters (orientation, gripper width) for grasps centered at each pixel in the input. \[\mathcal{M}: (\mathbf{I_d},\mathbf{I_s}) \mapsto \textbf{G}.\] To achieve this goal, we propose to use the [dnn]{acronym-label="dnn" acronym-form="singular+short"} in to map from depth and stiffness images to grasps **G** in the image, which we can easily transform to the real world using known coordinate transforms.
# Dataset {#sec:dataset}
To train *Def-GG-CNN* The pipeline of generating training data is visualized in . We first convert the triangular mesh of an object into tetrahedral mesh using fTetWild and feed that tetrahedral mesh to the Isaac Gym simulator to enable its soft bodies simulation feature. The stiffness of an object can then be varied by adjusting the material parameters, *i.e.*, Young's modulus and Poisson's Ratio. Then using the object triangular mesh, we sample grasp candidates which are later evaluated with Isaac Gym using proper quality metrics. Based on the performance of the grasps, we then label the grasps, convert them to the desired representation, and store them in the training dataset. **Depth and stiffness input:** We captured depth images of target objects with a virtual camera set to view the scene from top-down. To model variable object stiffness, four values of Young's modulus from \(2\cdot 10^4\) to \(2\cdot 10^9\) were used. The Young's modulus is normalized to \[0,1\] range and the corresponding stiffness value is assigned to every pixel in the stiffness image that the object occupies. **Grasp candidates:** Grasps are sampled with an antipodal grasp sampler to obtain approximately 200 grasp candidates for each target object. All grasps that collide with the mesh are filtered out, resulting in a final set of 25 to 40 collision-free grasps for each object. **Quality metrics:** None of the standard grasp quality metrics, such as the Ferrari & Canny \(L_1\) metric, are directly applicable for both rigid and deformable objects. As a quality metric, we use a *shake task* **Ground-truth grasp map:** To further simplify the data generation, we only use positively labeled grasps as ground-truth grasps to train the network. To generate the ground-truth grasps, we first transform all grasps to the image space. To do so, we represent the grasps as rectangles in the image as displayed in . Four parameters define the rectangles: grasp center, grasp orientation, grasp width, and finger height. Finally, we use the rectangles as image masks to generate ground-truth grasp maps **G**. Specifically, all pixels of the quality images **Q**, angle images \(\boldsymbol\phi\), and width images **W** within the rectangle are set to the values given from the *shake task* **Training dataset:** As a training dataset, we generate and label grasps on 30 objects. The objects include 13 primitive objects provided in Isaac Gym, 5 objects from the YCB dataset, and 12 objects with adversarial geometry from the EGAD! dataset. Because we set the stiffness for each object to four different values, the training set contains, in total, 120 objects. We use the Franka Emika Panda | {'timestamp': '2021-09-14T02:14:40', 'yymm': '2109', 'arxiv_id': '2109.05320', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05320'} |
null | null |
null | null |
# Introduction
Recently, three of us, together with D. J. Yang, have proposed a novel mechanism for generating transverse single-spin asymmetry (SSA) in semi-inclusive deep inelastic scattering (SIDIS) \(ep^\uparrow\to e'hX\). It has been demonstrated that an imaginary phase necessary for SSA can be produced purely within a parton-level cross section starting at two loops. The spin-dependent part of the cross section at high transverse momentum \(P_{hT}>1\) GeV (measured with respect to the virtual photon direction) can be schematically written as \[\frac{d\Delta \sigma}{dP_{hT}} \sim g_T(x)\otimes H\otimes D_1(z)+\cdots, \label{sche}\] where \(g_T(x)\) is the twist-three parton distribution function (PDF) associated with a transversely polarized proton, \(D_1\) is the unpolarized twist-two fragmentation function (FF) for the observed hadron \(h\), and \(H\) is the hard kernel starting at \({\cal O}(\alpha_s^2)\) (see also an earlier related work ). The terms omitted in ([\[sche\]](#sche){reference-type="ref" reference="sche"}) are proportional to the 'genuine twist-three' quark-gluon correlation functions \(\sim \langle \bar{\psi}gF\psi\rangle\) commonly called the Efremov-Teryaev-Qiu-Sterman (ETQS) functions. As is well known, the \(g_T\) distribution can be written as the sum of the Wandzura-Wilczek (WW) part and the genuine twist-three part \[g_T(x) = \int_x^1 \frac{dx'}{x'} \Delta q(x') + ({\rm genuine\ twist\ three})\,, \label{ww}\] where \(\Delta q(x)\) is the twist-two polarized (helicity) quark PDF. It is a consistent truncation of the result in Ref. to keep only the WW part in ([\[ww\]](#ww){reference-type="ref" reference="ww"}). The new source of SSA can then be entirely expressed in terms of the twist-two PDFs \(\Delta q(x)\) and the twist-two FFs \(D_1(z)\). This is a remarkable observation in striking contrast to the prevailing view that SSA at high-\(P_{hT}\) is explained by the ETQS functions and certain twist-three fragmentation functions (see a review and references therein). Unlike these higher-twist distributions, twist-two distributions are very well constrained by global QCD analyses. Therefore, the mechanism proposed in offers a unique part of SSA that can be unambiguously calculated without introducing free parameters. Moreover, in the transverse-momentum-dependent (TMD) PDF framework valid in the low-\(P_{hT}\) region (\(P_{hT}\lesssim 1\) GeV), a new source of SSA proportional to the \(g_{1T}(x,k_\perp)\) distribution (the TMD version of \(g_T(x)\)) has been identified, along with more than a dozen of new contributions involving various twist-three TMDs and FFs and hard kernels up to two loops. Again, this calls into question the prevailing view in the community (see, e.g., ) that SSA at low-\(P_{hT}\) is entirely attributed to the Sivers and Collins functions. The purpose of this paper is twofold. First, we extend the analysis of to gluon-initiated channels. There exists a gluonic counterpart of \(g_T(x)\), the twist-three \({\cal G}_{3T}(x)\) distribution for a transversely polarized proton. Its WW part is related to the twist-two polarized gluon PDF \(\Delta G(x)\). In complete analogy to ([\[sche\]](#sche){reference-type="ref" reference="sche"}), we find the structure \[\frac{d\Delta \sigma}{dP_{h\perp}} \sim {\cal G}_{3T}(x)\otimes H_g\otimes D_1(z) \sim \Delta G(x)\otimes H_g\otimes D_1(z)\,, \label{newglue}\] which again consists only of twist-two distributions after the WW approximation. We shall identify the two-loop diagrams that go into the hard kernel \(H_g\) and study their gauge invariance and infrared safety. Equation [\[newglue\]](#newglue){reference-type="eqref" reference="newglue"} is a novel gluon-initiated source of SSA in SIDIS to be considered jointly with the previously known mechanism which involves genuine twist-three, three-gluon correlators \(\langle FgFF\rangle\). Second, we perform a detailed numerical analysis of SSA and make predictions for its measurements at the future Electron-Ion Collider (EIC). In doing so, we neglect the 'usual' contributions from the ETQS and twist-three FFs, which have been intensively anallyzed, and focus on the new contributions in order to explore their importance clearly. The results can be viewed as a baseline for future EIC measurements of SSA. Deviations from our predictions, if observed, may be attributed to genuinely twist-three effects. This paper is organized as follows. In Section II, we describe the SIDIS setup and introduce kinematic variables. In Section III, we first review the result of obtained for the quark-initiated channel, and then propose an analogous, but novel contribution to SSA in the gluon-initiated channel. In Section IV, we perform a detailed analysis of the two-loop diagrams and calculate the hard coefficients in all the partonic channels. In Section V, we implement the obtained formulas numerically and make predictions according to the kinematic coverage of the EIC. We also present results relevant to the COMPASS experiments. Finally, we discuss our findings and conclude in Section VI. Appendices are devoted to a technical proof of the infrared finiteness of the factorization formulas at two-loop level.
# SIDIS Kinematics {#sec:kin}
In this section we give a brief review of polarized SIDIS \(e(l)p(P)\to e(l')h(P_h)X\) and introduce involved kinematic variables. We have in mind light hadron production specifically for \(h=\pi^\pm\). Heavy-quark production will be studied in a separate work. The spin-dependent part of the differential cross section is given by \[d^6\Delta\sigma = \frac{1}{2S_{ep}}\frac{d^3P_h}{(2\pi)^32E_h}\frac{d^3 l'}{(2\pi)^32E_{l'}} \frac{e^4}{(Q^2)^2} L^{\mu\nu}W_{\mu\nu}\,, \label{eq:xspol}\] where \(S_{ep}\equiv (l+P)^2\), \(Q^2\equiv-q^2=-(l-l')^2\), \(L^{\mu\nu}=2(l^\mu l'^\nu + l^\nu l'^\mu)-g^{\mu\nu}Q^2\) is the leptonic tensor, \(W^{\mu\nu}\) is the hadronic tensor, and \(\nu\) and \(\mu\) are the polarization indices of the virtual photon in the amplitude and the complex-conjugate amplitude, respectively. The Bjorken variable is denoted as \(x_B=Q^2/(2P\cdot q)\). We shall work in the so-called hadron frame, where the virtual photon and the proton move in the \(z\) direction with \[q^\mu=(0,0,0,-Q)\,, \qquad P^\mu=\left(\frac{Q}{2x_B},0,0, \frac{Q}{2x_B} \right)\.\] The incoming and outgoing leptons have the momenta \[l^\mu = \frac{Q}{2} (\cosh\psi ,\sinh\psi \cos \phi, \sinh\psi \sin \phi,-1)\,, \qquad l'^\mu = \frac{Q}{2} (\cosh\psi ,\sinh\psi \cos \phi, \sinh\psi \sin \phi, 1)\,, \label{eq:use}\] where \(\phi\) is the azimuthal angle relative to the \(z\) axis, and \[\cosh\psi \equiv \frac{2x_BS_{ep}}{Q^2}-1\.\] We adopt the standard variables \[y=\frac{P\cdot q}{P\cdot l}\,, \qquad z_f =\frac{ P\cdot P_h}{ P \cdot q}\,,\] with the relation \(x_ByS_{ep}=Q^2\). Another common variable is \(q_T = \sqrt{-q_t^2}\) where \[q^\mu_t \equiv q^\mu-\frac{P_h \cdot q}{P_h\cdot P} P^\mu-\frac{P \cdot q}{P \cdot P_h} P_h^\mu \.\] In the present frame, the transverse part of \(q^\mu_t\) reads \(\vec{q}_{tT} =-\vec{P}_{hT}/z_f\). The momentum of the final state hadron can then be parametrized as \[P_h^\mu = \frac{z_f Q}{2}\left(1 + \frac{q_T^2}{Q^2}, \frac{2 q_T}{Q} \cos\chi, \frac{2 q_T}{Q} \sin \chi,-1 + \frac{q_T^2}{Q^2}\right)\. \label{eq:Ph}\] For the transverse spin of the incoming proton we choose \[S_T^\mu = (0,\cos\Phi_S,\sin\Phi_S,0)\.\] In terms of the above variables, the differential cross section [\[eq:xspol\]](#eq:xspol){reference-type="eqref" reference="eq:xspol"} takes the following Lorentz invariant form \[\frac{d^6\Delta \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} = \frac{\alpha_{\rm em}^2}{128 \pi^4 x_B^2 S_{ep}^2 Q^2} z_f L_{\mu\nu} W^{\mu\nu}\. \label{eq:xspol2}\] In practice, instead of \(\phi\) and \(\chi\), it is more convenient to define the hadron and spin angles relative to the lepton plane, \[\begin{aligned}
\phi_h\equiv \phi-\chi\,, \qquad \phi_S\equiv \phi-\Phi_S\,, \end{aligned}\] in accordance with the Trento conventions. The cross section is then a function of \(\phi_h\) and \(\Phi_S-\chi = \phi_h-\phi_S\). The dependence on \(\phi_h\) can be factored out by decomposing the hadron tensor \(W^{\mu\nu}\) using the following set of vectors, \[\begin{split} & T^\mu = \frac{1}{Q}\left(q^\mu + 2 x_B P^\mu\right)\,,\\ & X^\mu = \frac{1}{q_T}\left[\frac{P_h^\mu}{z_f}-q^\mu-\left(1 + \frac{q_T^2}{Q^2}\right) x_B P^\mu\right]\,,\\ & Y^\mu = \epsilon^{\mu\nu\rho\sigma} Z_\nu X_\rho T_\sigma\,,\\ & Z^\mu =-\frac{q^\mu}{Q}\,, \end{split}\] which form nine independent tensors, \(\mathcal{V}_k^{\mu\nu}\) (see for explicit expressions), and their inverses, \(\tilde{\mathcal{V}}_k^{\mu\nu}\). Among them, the following six symmetric tensors contribute to the decomposition of \(W^{\mu\nu}\), \[\begin{split} & \tilde{\mathcal{V}}_1^{\mu\nu} = \frac{1}{2}\left(2 T^\mu T^\nu + X^\mu X^\nu + Y^\mu Y^\nu\right)\,,\\ &\tilde{\mathcal{V}}_2^{\mu\nu} = T^\mu T^\nu\,,\\ &\tilde{\mathcal{V}}_3^{\mu\nu} = -\frac{1}{2}(T^\mu X^\nu + X^\mu T^\nu)\,,\\ &\tilde{\mathcal{V}}_4^{\mu\nu} = \frac{1}{2}(X^\mu X^\nu-Y^\mu Y^\nu)\,,\\ &\tilde{\mathcal{V}}_8^{\mu\nu} = -\frac{1}{2}(T^\mu Y^\nu + Y^\mu T^\nu)\,,\\ &\tilde{\mathcal{V}}_9^{\mu\nu} = \frac{1}{2}(X^\mu Y^\nu + Y^\mu X^\nu)\.\\ \end{split} \label{eq:basis}\] With these tensors we can write \[L_{\mu\nu} W^{\mu\nu} = Q^2 \sum_{k = 1,\dots,4, 8,9} \mathcal{A}_k(\phi_h) [W_{\rho \sigma} \tilde{\mathcal{V}}_k^{\rho \sigma} ]\,, \label{eq:lw}\] where \[\mathcal{A}_k(\phi_h) = L_{\mu\nu} \mathcal{V}_k^{\mu\nu}/Q^2\,,\] have the explicit expressions \[\begin{split} &\mathcal{A}_1(\phi_h) = 1 + \cosh^2\psi \,,\\ &\mathcal{A}_2(\phi_h) =-2 \,,\\ &\mathcal{A}_3(\phi_h) =-\cos\phi_h \sinh 2\psi \,,\\ &\mathcal{A}_4(\phi_h) = \cos 2\phi_h \sinh^2 \psi \,,\\ &\mathcal{A}_8(\phi_h) =-\sin\phi_h \sinh 2\psi\,,\\ &\mathcal{A}_9(\phi_h) = \sin 2\phi_h \sinh^2 \psi\. \end{split} \label{eq:Ak}\] We are thus led to the representation (see for example ) \[\begin{split} \frac{d^6\Delta \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} & = \sin(\phi_h-\phi_S)\left(\mathcal{F}_1 + \mathcal{F}_2\cos\phi_h + \mathcal{F}_3 \cos 2\phi_h\right) + \cos(\phi_h-\phi_S)\left(\mathcal{F}_4\sin\phi_h + \mathcal{F}_5 \sin 2\phi_h\right)\\ & =\Big[F^{\sin(\phi_h-\phi_S)} \sin(\phi_h-\phi_S) + F^{\sin(2\phi_h-\phi_S)}\sin(2\phi_h-\phi_S) + F^{\sin\phi_S} \sin \phi_S\\ & + F^{\sin(3\phi_h-\phi_S)} \sin(3\phi_h-\phi_S) + F^{\sin(\phi_h + \phi_S)} \sin(\phi_h + \phi_S)\Big]\,,\\ \end{split} \label{fourier}\] with \[\begin{split} & F^{\sin(\phi_h-\phi_S)} = \mathcal{F}_1\,,\\ & F^{\sin(2\phi_h-\phi_S)} = \frac{\mathcal{F}_2 + \mathcal{F}_4}{2}\,,\\ & F^{\sin \phi_S} = \frac{-\mathcal{F}_2 + \mathcal{F}_4}{2}\,,\\ & F^{\sin(3\phi_h-\phi_S)} = \frac{\mathcal{F}_3 + \mathcal{F}_5}{2}\,,\\ & F^{\sin(\phi_h + \phi_S)} = \frac{-\mathcal{F}_3 + \mathcal{F}_5}{2}\.\\ \end{split} \label{FF}\] The Fourier components \(\sin (\phi_h-\phi_S)\) and \(\sin (\phi_h+\phi_S)\) are referred to as the Sivers and Collins asymmetries, respectively. While we continue to use this nomenclatures, we emphasize that the new mechanism, which contributes to these asymmetries and will be studied in detail below, has nothing to do with the Sivers and Collins functions, or their collinear twist-three counterparts.
# New contributions to SSA
In this section we first recapitulate the \(g_T(x)\) contribution to SSA discussed in, and apply the so-called Wandzura-Wilczek (WW) approximation to simplify the result. We then derive another new contribution to SSA due to the gluonic counterpart of \(g_T\).
## Quark-initiated channel
In, it has been shown that the imaginary phase necessary for SSA in SIDIS can come from the hard kernel in perturbation theory starting at two loops, and all the relevant two-loop diagrams have been identified. However, only the quark (and antiquark) initiated channel was considered there. In this channel, motivated by the structure ([\[sche\]](#sche){reference-type="ref" reference="sche"}), we factorize the fragmentation function out of the hadronic tensor \(W_{\mu\nu}\) as \[W_{\mu\nu} = \sum_{a = q,\bar{q},g}\int \frac{dz}{z^2} D_1^a(z) w^a_{\mu\nu}\,, \label{eq:WDfact}\] where we have taken into account the fact that the observed hadron can also come from the fragmentation of a radiated gluon in the final state. The result of reads (suppressing the label \(a\) for simplicity) \[\begin{split} w_{\mu\nu} & = \frac{M_N}{2}\int dx g_T(x) {\rm Tr} \left[\gamma_5 \slashed{S}_T S^{(0)}_{\mu\nu}(x P) \right]\\ &-\frac{M_N}{4}\int dx \tilde{g}(x) {\rm Tr}\left[ \gamma_5\slashed{P} S_T^\alpha \left.\frac{\partial S^{(0)}_{\mu\nu}(k)}{\partial k_T^\alpha}\right|_{k=xP} \right]\\ &+ \frac{iM_N}{4}\int dx_1 dx_2 {\rm Tr}\left[\left(\slashed{P} \epsilon^{\alpha P n S_T} \frac{G_F(x_1,x_2)}{x_1-x_2} + i\gamma_5\slashed{P} S_T^\alpha \frac{\tilde{G}_F(x_1,x_2)}{x_1-x_2} \right) S_{\mu\nu\alpha}^{(1)}(x_1 P,x_2 P) \right]\,, \end{split} \label{eq:wtot}\] in which \(M_N\) is the proton mass, and \({\rm Tr}\) denotes trace over colors and Dirac indices. Our conventions are \(\epsilon_{0123}=+1\), \(\gamma_5=i\gamma^0\gamma^1\gamma^2\gamma^3\) and \(\epsilon^{\alpha PnS} =\epsilon^{\alpha\beta\rho\lambda}P_\beta n_\rho S_{T\lambda}\) with the light-like vector \(n^\mu\) satisfying \(n^2 = 1\) and \(n\cdot P = 1\). The \(g_T(x)\) distribution function is defined as \[\int \frac{d\lambda}{2\pi} e^{i\lambda x} \langle P S_T |\bar{\psi}_j(0)[0,\lambda n]\psi_i(\lambda n)| P S_T \rangle = \frac{M_N}{2}(\gamma_5\slashed{S}_T)_{ij} g_T(x) + \dots\. \label{gtdef}\] while \(G_F(x_1,x_2)\) and \(\tilde{G}_F(x_1,x_2)\) are the ETQS functions (We follow the notation of Ref. where explicit definitions can be found). The 'kinematical' distributions \(\tilde{g}(x)\) and \(g_T(x)\) are related through the QCD equation of motion \[g_T(x) + \frac{\tilde{g}(x)}{2x} = \int d x' \frac{G_F(x,x') + \tilde{G}_F(x,x')}{x-x'} ~. \label{eq:gtgtil}\] The hard matrix elements \(S_{\mu\nu}^{(0)}(xP)\) and \(S^{(1)}_{\mu\nu\alpha}(x_1P,x_2P)\) are computable in perturbation theory. As observed in, the first nonzero contribution to \(S^{(0)}\) appears at two loops, \(S^{(0)}\propto \alpha_s^2\), and \(S^{(1)}\) is obtained from \(S^{(0)}\) by attaching an extra gluon in all possible ways. Two representative diagrams contributing to \(S^{(0)}\) are shown in Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}. The crosses denote on-shell lines that lead to an imaginary phase via the Cutkosky rules. We note that Ref. arrived at essentially the same structure as ([\[eq:wtot\]](#eq:wtot){reference-type="ref" reference="eq:wtot"}), but did not specify the hard kernels \(S^{(0,1)}\). We shall compute \(w_{\mu\nu}\) in the WW approximation, namely, by systematically neglecting genuine twist-three distributions everywhere. This is a consistent approximation in the sense that it preserves both QED and QCD gauge invariance. In this approximation, we may write[^1] \[g_T(x) \to \int_x^1 \frac{dx'}{x'} (\Delta q(x')+\Delta\bar{q}(x))\,, \qquad \tilde{g}(x) \approx-2xg_T(x)\,, \label{eq:gTww}\] where \(\Delta q(x)\) and \(\Delta\bar{q}(x)\) are the standard twist-two polarized quark and antiquark distributions. Moreover, the first two lines in [\[eq:wtot\]](#eq:wtot){reference-type="eqref" reference="eq:wtot"} can be combined into \[w_{\mu\nu} \approx \frac{M_N}{2}\int dx g_T(x)S_T^\alpha\left( \frac{\partial}{\partial k_T^\alpha}{\rm Tr}[\gamma_5 \slashed{k} S_{\mu\nu}^{(0)}(k)]\right)_{k = x P}\. \label{eq:wsmalww}\] We thus arrive at the structure mentioned in the introduction, \[\begin{aligned}
d\Delta \sigma \sim (\Delta q(x)+\Delta \bar{q}(x)) \otimes H \otimes D_1(z)\. \label{above} \end{aligned}\] The above formulas hold for each quark flavor. In practice, we must sum over flavors weighted by the quark electromagnetic charge. In physical cross sections, we thus apply \(g_T(x) \to \sum_f e_f^2 g_{Tf}(x)\) where \(g_{Tf}\) is given by ([\[eq:gTww\]](#eq:gTww){reference-type="ref" reference="eq:gTww"}) for each quark flavor \(f\). Let us compare ([\[above\]](#above){reference-type="ref" reference="above"}) with the conventional contribution from the ETQS function which schematically reads[^2] \[\begin{aligned}
d\Delta \sigma_{\rm ETQS} \sim G_{F} (x,x')\otimes H'\otimes D_1(z)\. \label{etq} \end{aligned}\] Since \(H\sim {\cal O}(\alpha_s^2)\) and \(H'\sim {\cal O}(\alpha_s)\), naively the former is parametrically suppressed by a factor of \(\alpha_s\). However, the definition \(G_F\sim \langle \bar{\psi}gF\psi\rangle\) explicitly contains the coupling \(g\) which actually comes from perturbative diagrams. That is, some suppression associated with the coupling \(g\) goes into \(G_F\) in the convention ([\[etq\]](#etq){reference-type="ref" reference="etq"}). As for the soft part, both conceptually and practically, we have a far better grasp of twist-two distributions than twist-three distributions: \(\Delta q\) and \(D_1\) have been well constrained thanks to a wealth of experimental data and global QCD analyses, whereas the ETQS functions are still poorly constrained. These considerations make ([\[above\]](#above){reference-type="ref" reference="above"}) a new and attractive source of SSA that can be unambiguously calculated without introducing any free parameters. The main goal of this paper is to carry out such a calculation, both analytically and numerically. But before doing so, let us point out that an entirely analogous contribution exists in the gluon-initiated channel.
## Gluon-initiated channel
The gluonic counterpart of \(g_T(x)\) for a transversely polarized proton is defined as \[\int \frac{d\lambda}{2\pi} e^{ix\lambda}\langle PS_T |F^{n\alpha}(0)[0,\lambda n] F^{n\beta}(\lambda n)| PS_T\rangle = i M_N x\mathcal{G}_{3T}(x) \epsilon^{n \alpha\beta S_T} + \dots\. \label{eq:G3T}\] Similar to \(g_T(x)\), it can be written as the sum of the WW part and the genuine twist-three part, \[{\cal G}_{3T}(x) =\frac{1}{2} \int_x^1 \frac{dx'}{x'} \Delta G(x')+({\rm genuine\ twist\ three})\,, \label{eq:G3Tww}\] where the WW part is related to the polarized (helicity) gluon PDF \(\Delta G(x)\), and the genuine twist-three part consists of three-gluon correlators \(\sim \langle FFF\rangle\). Their full expressions can be found in. The \({\cal G}_{3T}\) distribution appeared in the previous calculation of the double spin asymmetry \(A_{LT}\) in proton-proton collisions \(p^\rightarrow p^\uparrow \to hX\) . The cross section formula derived in can be straightforwardly adapted to the case of single spin asymmetry in SIDIS \(e p^\uparrow \to hX\). Writing the hadronic tensor as \[\begin{aligned}
W_{\mu\nu}^g=\sum_{a=q,\bar{q}}\int \frac{dz}{z^2}D_1^a(z)w_{\mu\nu}^{g,a}\,, \end{aligned}\] we find (see (17) and (25) of ) \[\begin{split} w^g_{\mu\nu} =& i M_N \int \frac{dx}{x} \mathcal{G}_{3T}(x)\epsilon^{n\alpha\beta S_T} S_{\mu\nu}^{(0) \alpha' \beta'} (xP)\omega_{\alpha'\alpha} \omega_{\beta'\beta}\\ &-i M_N \int \frac{dx}{x^2} \tilde{g}(x)\left(g_T^{\beta\lambda} \epsilon^{\alpha P n S_T}-g_T^{\alpha\lambda} \epsilon^{\beta P n S_T}\right)\left( \frac{\partial S_{\mu\nu\alpha \beta}^{(0)} (k)}{\partial k^\lambda}\right)_{k = xP} \\ &-\frac{1}{2}\int \frac{dx_1dx_2}{x_1x_2} M_F^{\alpha\beta\gamma}(x_1,x_2)\frac{S^{(1)\alpha'\beta'\gamma'}(x_1,x_2)}{x_2-x_1}\omega_{\alpha'\alpha}\omega_{\beta'\beta}\omega_{\gamma'\gamma} \,, \end{split} \label{eq:wglue}\] where \(g_T^{\mu\nu} = g^{\mu \nu}-P^\mu n^\nu-n^\mu P^\nu\) and \(\omega^{\mu\nu} = g^{\mu\nu}-P^\mu n^\nu\) are the projectors to the transverse space. \(\tilde{g}(x)\) is again a kinematical function with its precise definition given in (see also where it is called \(\Delta G^{(1)}_T(x)\)). \(M_F\) denotes the three-gluon correlators \(\langle FFF\rangle\) (see (19) of ). The hard part also starts at two loops, \(S^{(0)}\sim \alpha_s^2\), whose diagrams have the same topology as in the quark-initiated channel. A representative diagram is displayed on the left hand side of Fig. [\[fig:gluon\]](#fig:gluon){reference-type="ref" reference="fig:gluon"}, and \(S^{(1)}\) is obtained by attaching a gluon to this diagram in all possible ways. The diagram on the right, which is an analog of the right diagram in Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}, does not contribute due to Furry's theorem. Note that in the computation of \(A_{LT}\) in Ref. , the imaginary phase comes from the definition of \(\Delta G(x)\) for the longitudinally polarized proton. That is, the non-pole part of the hard kernel was calculated. In the present case, the imaginary phase comes from propagator poles in the hard kernel \(S^{\mu\nu}_{\alpha\beta}\), and this is why two-loop diagrams are needed. In the WW approximation, we may write \[{\cal G}_{3T}(x) \approx \frac{1}{2} \int_x^1 \frac{dx'}{x'} \Delta G(x')\,, \qquad \tilde{g}(x) \approx x^2 \mathcal{G}_{3T}(x)\,,\] and neglect the third line of ([\[eq:wglue\]](#eq:wglue){reference-type="ref" reference="eq:wglue"}). We thus arrive at a new contribution to SSA of the form ([\[newglue\]](#newglue){reference-type="ref" reference="newglue"}) which consists only of twist-two distributions. For light-hadron production, this contribution is suppressed compared to the quark one discussed earlier. However, for SSA in productions of heavy systems such as open charm and \(J/\psi\), it is expected to play a more important role.
# Computation of the hard part: Quark-initiated channel
In this and the next sections, we embark on an analysis of the two-loop diagrams for the quark and gluon initiated channels, respectively. The calculation is rather involved, especially because nontrivial cancellations of infrared divergences are in demand. In the end, we shall have infrared safe formulas that can be straightforwardly evaluated numerically.
## Quark-fragmenting channel
In Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}, either a quark or a gluon in the final state fragments into the observed hadron. For definiteness, we focus on the former process below. The treatment of the latter is basically analogous, and will be included only in the final formulas. The hard factor \(S^{(0)}_{\mu\nu}\) for the quark-initiated and quark-fragmenting channel explicitly reads \[\begin{split} S^{(0)\mu \nu}(k) & =-\frac{g^4}{2N_c}(2\pi)\delta\left(\left(k+q-p_q\right)^2\right)\int \frac{d^4 l_2}{(2\pi)^4} (2\pi)\delta\left(l_2^2\right)(2\pi) \delta\left((k + q-l_2)^2\right)\\ &\times \left\{ i A^{\alpha\mu}(k+q-p_q) \bar{M}_{\alpha\beta}(k+q-p_q,l_2)A^{\nu\beta}(l_2)-iA^{\alpha\mu}(l_2) \bar{M}_{\alpha\beta}(l_2,k+q-p_q) A^{\nu\beta}(k+q-p_q)\right\}\,, \label{eq:S0} \end{split}\] where \(k\) and \(p_q\) are the the momenta of incoming and outgoing quarks, respectively, and \[\begin{aligned}
&&\bar{M}_{\alpha\beta}(k+q-p_q,l_2)\nonumber \\ && = \frac{N_c(N_c^2-1)}{4}\slashed{p}_q\left[-\frac{V_{\alpha\beta\rho}(k+q-p_q,l_2)\gamma^\rho}{(k+q-p_q-l_2)^2} + \frac{N_c^2-1}{N_c^2}\gamma_\alpha \frac{\slashed{k} + \slashed{q}}{(k+ q)^2}\gamma_\beta-\frac{1}{N_c^2}\gamma_\beta \frac{\slashed{p}_q-\slashed{l}_2}{(p_q-l_2)^2}\gamma_\alpha\right](\slashed{k}+\slashed{q}-\slashed{l}_2)\,, \label{eq:mm} \end{aligned}\] \[V_{\alpha\beta\rho}(k+q-p_q,l_2) = g_{\alpha\beta}(l_2+k+q-p_q)_\rho + g_{\alpha\rho}(l_2-2(k+q-p_q))_\beta + g_{\rho\beta}(k+q-p_q-2l_2)_\alpha\,,\] \[A^{\alpha\mu}(k+q-p_q)= \gamma^\alpha \frac{\slashed{p}_q-\slashed{q}}{(p_q-q)^2}\gamma^\mu +\gamma^\mu\frac{\slashed{k} + \slashed{q}}{(k + q)^2}\gamma^\alpha\,, \label{eq:Al1}\] \[A^{\nu\beta}(l_2)= \gamma^\nu \frac{\slashed{k}-\slashed{l}_2}{(k-l_2)^2}\gamma^\beta +\gamma^\beta\frac{\slashed{k} + \slashed{q}}{(k + q)^2}\gamma^\nu\. \label{eq:Al2}\] Equation ([\[eq:S0\]](#eq:S0){reference-type="ref" reference="eq:S0"}) represents the sum of \(12=2\times 3\times 2\) diagrams, two of which are shown in Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}. One can easily recognize the part of diagrams each piece of Feynman rules corresponds to. In [\[eq:mm\]](#eq:mm){reference-type="eqref" reference="eq:mm"} we have performed a color trace, while the Dirac trace is yet to be done. There are three \(\delta\)-functions, one for the unobserved gluon in the final state \(\delta((k+q-p_q)^2)\), and the other two come from the poles of internal propagators (denoted by the crosses in Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}). Eventually we shall take the collinear limit \(k \to p_1 \equiv xP\) in these expressions and introduce shorthand notations \(p_2 \equiv p_1 + q\) and \(l_1 \equiv p_2-p_q\), the latter being the momentum of the unobserved gluon in Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}. However, this has to be done with some care because the limit does not commute with the \(k_{T}\)-derivative acting on the hard kernel in ([\[eq:wsmalww\]](#eq:wsmalww){reference-type="ref" reference="eq:wsmalww"}). Let us define \[S^{(0)}_{\mu\nu}(k) = g^4 (2\pi) \delta\left(\left(k+q-p_q\right)^2\right)\int \frac{d^2 \vec{l}_{2T} dl_2^+}{(2\pi)^3 2l_2^+}(2\pi) \delta\left((k + q-l_2)^2\right) \widehat{S}^{(0)}_{\mu\nu}(k)\,, \label{eq:S0k}\] where \(\widehat{S}^{(0)}_{\mu\nu}(k)\) can be read off from [\[eq:S0\]](#eq:S0){reference-type="eqref" reference="eq:S0"}. We first convert the \(k_T\)-derivatives of the \(\delta\)-functions to the \(x\)-derivatives as \[\begin{aligned}
&& S_T^\alpha\left(\frac{\partial}{\partial k_T^\alpha} \delta\left(\left(k + q-p_q\right)^2\right)\right)_{k = p_1} =-\frac{p_q \cdot S_T}{p_1 \cdot (p_2-p_q)} x\frac{\partial}{\partial x}\delta\left((p_2-p_q)^2\right) = \frac{l_1\cdot S_T}{p_1\cdot l_1}x\frac{\partial}{\partial x}\delta(l_1^2)\,, \label{del1}\\ && S_T^\alpha\left(\frac{\partial}{\partial k_T^\alpha} \delta\left(\left(k + q-l_2\right)^2\right)\right)_{k = p_1} =-\frac{l_2 \cdot S_T}{p_1 \cdot (p_2-l_2)} x\frac{\partial}{\partial x}\delta\left((p_2-l_2)^2\right)\,, \label{del2} \end{aligned}\] and then use integration by parts to shuffle the \(x\)-derivatives from the \(\delta\)-functions to the hard factor \(\widehat{S}^{(0)}_{\mu\nu}\). From the term \(\partial\delta(l_1^2)/\partial x\) in ([\[del1\]](#del1){reference-type="ref" reference="del1"}), we get a term with \(\partial g_T/\partial x\) and a term with \[\frac{\partial}{\partial x}\left[x\int \frac{d^2 \vec{l}_{2T} dl_2^+}{(2\pi)^3 2l_2^+}\dots\right]\. \label{eq:loopx}\] From the term \(\partial\delta\left((p_2-l_2)^2\right)/\partial x\) in ([\[del2\]](#del2){reference-type="ref" reference="del2"}), we get a term with \(\partial\delta(l_1^2)/\partial x\), a term with \(\partial g_T/\partial x\) and a term with \[\frac{\partial}{\partial x}\left\{{\rm Tr}\left[\gamma_5 \slashed{p}_1\widehat{S}^{(0)}_{\mu\nu}(p_1)\right]\right\}\.\] We further convert \(\partial \delta(l_1^2)/\partial x\) to a term with \(\partial g_T/\partial x\) and a term like [\[eq:loopx\]](#eq:loopx){reference-type="eqref" reference="eq:loopx"}. The two resulting terms with \(\partial g_T / \partial x\) cancel. In total, we are led to \[\begin{aligned}
w_{\mu\nu} &=& \frac{M_N}{2}\int dx g_T(x) S_T^\alpha\left(\frac{\partial}{\partial k_T^\alpha}{\rm Tr}[\gamma_5 \slashed{k} S^{(0)}_{\mu\nu}(k)]\right)_{k = p_1} \nonumber \\ &=& \frac{M_N}{2}\int dx (2\pi)\delta\left(l_1^2\right) \Biggl[-x\frac{\partial g_T(x)}{\partial x} \int \frac{d^2\vec{l}_{2T} dl_2^+}{(2\pi)^3 2l_2^+}\frac{l_1\cdot S_T}{p_1\cdot l_1} (2\pi) \delta\left((p_2-l_2)^2\right) g^4 {\rm Tr}[\gamma_5 \slashed{p}_1 \widehat{S}^{(0)}_{\mu\nu}(p_1)]\nonumber \\ && \qquad \quad-g_T(x)\frac{\partial}{\partial x}\left\{ x \int \frac{d^2\vec{l}_{2T} dl_2^+}{(2\pi)^3 2l_2^+}\left[\frac{l_1\cdot S_T}{p_1\cdot l_1} + \frac{l_2\cdot S_T}{p_1\cdot (p_2-l_2)}\right] (2\pi) \delta\left((p_2-l_2)^2\right) g^4 {\rm Tr}\left[\gamma_5 \slashed{p}_1 \widehat{S}^{(0)}_{\mu\nu}(p_1)\right]\right\}\nonumber \\ && \qquad \quad +x g_T(x)\int \frac{d^2\vec{l}_{2T} dl_2^+}{(2\pi)^3 2l_2^+}\frac{l_2 \cdot S_T}{p_1 \cdot (p_2-l_2)} (2\pi)\delta\left((p_2-l_2)^2\right) g^4 \frac{\partial}{\partial x} \left\{{\rm Tr}[\gamma_5 \slashed{p}_1 \widehat{S}^{(0)}_{\mu\nu}(p_1)]\right\}\nonumber \\ && \qquad \quad + g_T(x) \int \frac{d^2\vec{l}_{2T} dl_2^+}{(2\pi)^3 2l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right) g^4 S_T^\alpha \left(\frac{\partial}{\partial k_T^\alpha}{\rm Tr}[\gamma_5 \slashed{k} \widehat{S}^{(0)}_{\mu\nu}(k)]\right)_{k = p_1}\Biggr]\,, \label{eq:lines} \end{aligned}\] for which \(l_1\) is fixed through momentum conservation as \(l_1 = p_2-p_q\). A general proof on both QED and QCD gauge invariance of the hadronic tensor ([\[eq:wtot\]](#eq:wtot){reference-type="ref" reference="eq:wtot"}) was given in. It was also realized that the first two terms in ([\[eq:wtot\]](#eq:wtot){reference-type="ref" reference="eq:wtot"}) contain infrared divergence separately when the momentum \(l_2\) becomes collinear to the incoming quark line (see Fig. [\[fig:quark\]](#fig:quark){reference-type="ref" reference="fig:quark"}), but the divergences cancel exactly. Now that we have written the original formula in a significantly different form ([\[eq:lines\]](#eq:lines){reference-type="ref" reference="eq:lines"}), it is a nontrivial task to check that ([\[eq:lines\]](#eq:lines){reference-type="ref" reference="eq:lines"}) is gauge invariant and divergence free. In Appendix A, we show that this is indeed the case, but only after summing all the lines of ([\[eq:lines\]](#eq:lines){reference-type="ref" reference="eq:lines"}). Knowing where divergences are hidden in intermediate expressions greatly helps a numerical analysis.
## Calculation of the hard coefficients
With [\[eq:lw\]](#eq:lw){reference-type="eqref" reference="eq:lw"}, [\[eq:WDfact\]](#eq:WDfact){reference-type="eqref" reference="eq:WDfact"} and [\[eq:lines\]](#eq:lines){reference-type="eqref" reference="eq:lines"}, the polarized cross section [\[eq:xspol2\]](#eq:xspol2){reference-type="eqref" reference="eq:xspol2"} takes the following form \[\begin{split} & \frac{d^6\Delta \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} = \frac{\alpha_{\rm em}^2 \alpha_S^2 M_N}{16\pi^2 x_B^2 S_{ep}^2 Q^2} \sum_k \mathcal{A}_k \int \frac{d x}{x} \int \frac{dz}{z} (2\pi)\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2 D_f(z) \\ &\qquad \times \Bigg\{-x^2\frac{\partial g_{T f}(x)}{\partial x} \frac{l_1 \cdot S_T}{p_1 \cdot l_1} \int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta((p_2-l_2)^2) {\rm Tr}\left[\gamma_5 \slashed{p}_1 \widehat{S}_{\mu\nu}^{(0)}(p_1)\tilde{\mathcal{V}}^{\mu\nu}_k\right]\\ & \qquad \qquad-x g_{Tf}(x) \frac{\partial}{\partial x}\Bigg\{x\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right)\left[ \frac{l_1 \cdot S_T}{p_1\cdot l_1} + \frac{l_2 \cdot S_T}{p_1\cdot (p_2-l_2)}\right] {\rm Tr}\left[\gamma_5 \slashed{p}_1 \widehat{S}_{\mu\nu}^{(0)}(p_1)\tilde{\mathcal{V}}^{\mu\nu}_k\right]\Bigg\}\\ &\qquad \qquad +xg_{Tf}(x)\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2l_2^+} (2\pi)\delta\left((p_2-l_2)^2\right) \frac{l_2 \cdot S_T}{p_1\cdot (p_2-l_2)} x\frac{\partial}{\partial x}\left\{{\rm Tr}\left[\gamma_5 \slashed{p}_1 \widehat{S}_{\mu\nu}^{(0)}(p_1)\tilde{\mathcal{V}}^{\mu\nu}_k\right]\right\}\\ &\qquad \qquad +xg_{Tf}(x)\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right)\left(\frac{\partial}{\partial k_T^\alpha}{\rm Tr}\left[\gamma_5 \slashed{k} S_T^\alpha \widehat{S}_{\mu\nu}^{(0)}(k)\tilde{\mathcal{V}}^{\mu\nu}_k\right]\right)_{k = p_1}\Bigg\}\,, \end{split} \label{eq:px}\] where the common notations \(\hat{x}\equiv x_B/x\) and \(\hat{z}\equiv z_f/z\) have been introduced. We have included a flavor summation with explicit charges \(e_f^2\), as commented after ([\[above\]](#above){reference-type="ref" reference="above"}). Equation ([\[eq:px\]](#eq:px){reference-type="ref" reference="eq:px"}) contains two \(\delta\)-function constraints with the first one \(\delta\left(\frac{q_T^2}{Q^2}-\cdots \right)\) originating from \(\delta\left(l_1^2\right)\). Solving the conditions \(l_1^2=(p_2-p_q)^2=0\) and \(p_q^2 = (p_2-l_1)^2 = 0\), we find two roots \[l_{1 (a_1)}^+ = \frac{p_2^+}{2}\left( 1 + a_1 \Delta_1\right) \,, \qquad l_{1 (a_1)}^-= \frac{p_2^-}{2}\left( 1-a_1 \Delta_1\right)\,, \qquad \Delta_1 = \sqrt{1-\frac{4 l_{1 T}^2}{p_2^2}} \,, \qquad a_1 = \pm 1\. \label{eq:pqpm}\] Recalling the definition \(z_f = P\cdot P_h/P\cdot q\), we have \(p_q^-= p_2^--l_1^-= \hatz q^-\), whose matching onto [\[eq:pqpm\]](#eq:pqpm){reference-type="eqref" reference="eq:pqpm"} leads to \(\Delta_1 = a_1(2\hatz-1)\). Since \(\Delta_1 > 0\), the two roots \(l_{1 (a_1)}^+\) effectively split the \(z\) integration according to the constraint \(a_1 (2 \hatz-1) > 0\): \[\int dz = \sum_{a_1 = \pm}\int dz \theta\left(a_1 (2\hatz-1)\right)\. \label{eq:dzbrok}\] The second \(\delta\)-function sets \((p_2-l_2)^2=0\), which, together with the condition \(l_2^2=0\), give two roots \[l_{2 (a_2)}^+ = \frac{p_2^+}{2}\left( 1 + a_2 \Delta_2\right) \,, \qquad l_{2 (a_2)}^-= \frac{p_2^-}{2}\left( 1-a_2 \Delta_2\right)\,, \qquad \Delta_2 = \sqrt{1-\frac{4 l_{2T}^2}{p_2^2}} \,,\qquad a_2 = \pm 1\. \label{eq:l2pm}\] We are allowed to perform the \(l_2\) integrals as \[\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right) = \sum_{a_2 = \pm} \int \frac{d^2 \vec{l}_{2T}}{(2\pi)^2}\frac{1}{2 p_2^2 \Delta_2}\int_{-\infty}^\infty d l_2^+ \delta\left(l_2^+-l_{2 (a_2)}^+\right) = \frac{1}{32 \pi^2}\sum_{a_2 = \pm} \int_0^1 d\Delta_2 \int_0^{2\pi} d\phi_2\. \label{eq:l2int}\] In the last equality, we have switched to the polar coordinate and changed the integration variable from \(l_{2 T}\) to \(\Delta_2\). This facilitates the computation significantly because we do not have to integrate over rational functions involving square roots. Next, we compute the Dirac traces using Feyncalc and apply the \(x\)-and \(k_T\)-derivatives to the 3rd, 4th and 5th lines of ([\[eq:px\]](#eq:px){reference-type="ref" reference="eq:px"}), which have to be done carefully. Note that the \(x\)-derivative acts outside the \(l_2\) integral in the 3rd line. We can only evaluate \(\delta\left((p_2-l_2)^2\right)\) and the evaluation of \(\delta\left(l_1^2\right)\) cannot be performed before taking the \(x\)-derivative. In the 4th (5th) line the \(x\) (\(k_T^\alpha\))-derivative is within the \(l_2\) integral and so both \(p_2-l_2\) and \(l_1\) are put on-shell after the derivatives are taken. The subsequent integrals over \(\phi_2\) and \(\Delta_2\) are the most cumbersome part of the entire calculation. The nontrivial angular dependence comes from the propagator denominators \[\begin{split} & (l_1-l_2)^2 =-\left[\frac{p_2^2}{4}(a_1\Delta_1-a_2\Delta_2)^2 + (\vec{l}_{1T}-\vec{l}_{2T})^2 \right]\,,\\ & (p_2-l_1-l_2)^2 =-\left[\frac{p_2^2}{4}(a_1\Delta_1 + a_2 \Delta_2)^2 + (\vec{l}_{1T} + \vec{l}_{2T})^2 \right]\,, \end{split}\] leading to a \(\cos(\phi_1-\phi_2)\) term, while in the numerator, after taking Dirac traces, we are left with powers of \(\cos(\phi_1-\phi_2)\) as well as linear terms of the forms \(\sin(\phi_2-\Phi_S)\) and \(\cos(\phi_2-\Phi_S)\) arising from \(\epsilon^{P n l_2 S_T}\) and \(l_2\cdot S_T\), respectively. We list the formulas used to carry out such integrals in Appendix [\[sec:integral\]](#sec:integral){reference-type="ref" reference="sec:integral"}. After the \(\phi_2\) integration, the \(k=1,2,3,4\) terms are proportional to \[\begin{aligned}
\epsilon^{l_{1T}S_T} =-\frac{1}{z}\epsilon^{P_{hT}S_T} =-q_T \hat{z} \sin(\Phi_S-\chi)\,, \end{aligned}\] where \(\epsilon^{12}=-\epsilon^{21}=1\), and the \(k=8,9\) terms are proportional to \[\begin{aligned}
l_1\cdot S_T =-\vec{l}_{1T}\cdot \vec{S}_T = \frac{1}{z}\vec{P}_{hT}\cdot \vec{S}_T = q_T\hat{z}\cos(\Phi_S-\chi)\. \end{aligned}\] In the individual lines of ([\[eq:px\]](#eq:px){reference-type="ref" reference="eq:px"}), the integral over the modulus \(l_{2T}\) has a singularity when \(l_{2T} \to 0\) (or when \(\Delta_2 \to 1\)), and when \(a_2 =-1\). However, as we will demonstrate in Appendix A, the total expression is finite because of the QCD Ward identity. Therefore, we first compute the \(\phi_2\) integrals for each line separately, sum up the results from all the lines and perform the \(l_{2T}\) (\(\Delta_2\)) integration afterwards. One notable feature is that the loop integration yields in principle a different expression for each of the four combinations of the roots \((a_1,a_2)\). However, we have found that after the summation over \(a_2\) the results are independent of \(a_1\). This is an important consistency check as it effectively ensures that, after all, the split [\[eq:dzbrok\]](#eq:dzbrok){reference-type="eqref" reference="eq:dzbrok"} is not necessary and we are back to the ordinary \(z\) integral over a complete domain allowed by kinematics. The above discussion is for the quark-initiated and quark-fragmenting channel. We have repeated the whole procedure for the quark-initiated and gluon-fragmenting channel. Adding the two pieces, we finally arrive at the total result \[\begin{split} & \frac{d^6\Delta \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} = \frac{\alpha_{\rm em}^2 \alpha_S^2 M_N}{16\pi^2 x_B^2 S_{ep}^2 Q^2} \sum_k \mathcal{A}_k \mathcal{S}_k \int^1_{x_{\rm min}} \frac{d x}{x} \int^1_{z_{\rm min}} \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \\ &\times\sum_f e_f^2 \big[D_f(z) x^2\frac{\partial g_{T f}(x)}{\partial x}\Delta\hat{\sigma}^{qq}_{D k} + D_f(z) x g_{Tf}(x) \Delta\hat{\sigma}^{qq}_k + D_g(z) x^2\frac{\partial g_{T f}(x)}{\partial x}\Delta\hat{\sigma}^{qg}_{D k} + D_g(z) x g_{Tf}(x) \Delta\hat{\sigma}^{qg}_k\big]\,, \end{split} \label{eq:ddsig}\] where \({\cal S}_k = \sin (\Phi_S-\chi)\) for \(k=1,2,3,4\) and \({\cal S}_k=\cos (\Phi_S-\chi)\) for \(k=8,9\). Note that we may substitute \(x\partial g_{Tf}/\partial x \approx -\Delta q_f(x)\) in the above expression. The hard coefficients in the quark-fragmenting (\(qq\)) channel are given by \[\begin{split} \Delta\hat{\sigma}_{D8}^{qq} &= \frac{\left(N_c^2-1\right) \hatx \hatz}{ 2N_c^2 Q {\left(1-\hatz\right)}^2}\Biggl[ \left(1-\hatz\right) \Bigl(1-\hatx+\hatz-3\hatx \hatz+N_c^2 \left(1-\hatx-\hatz+3\hatx \hatz\right)\Bigr)+2\left(1-2\hatx\right) \hatz \log(\hatz) \Biggr]\,, \\ \Delta\hat{\sigma}_{D9}^{qq} &=\frac{\left(N_c^2-1\right) \left(1-\hatx\right) \hatx \hatz }{2N_c^2 q_T {\left(1-\hatz\right)}^2} \Biggl[ \left(1-\hatz\right) \left(3N_c^2 \left(1-\hatz\right)+3\hatz-1\right)-2\left(1-2\hatz\right) \log(\hatz)\Biggr]\,, \end{split}\] \[\begin{split} \Delta\hat{\sigma}_1^{qq} &=\frac{N_c^2-1 }{2N_c^2 q_T \left(1-\hatz\right) \hatz} \Bigg[(1-\hatz)\bigl(\hatz\hatx \left(\hatx \left(3+10\hatz\right)-3\left(1+\hatz\right)\right)-1\bigr) \\ & \qquad \qquad +N_c^2 \hatz \left({\hatx}^2 \left(3+2\hatz \left(5\hatz-6\right)\right)-1-3\hatx {\left(1-\hatz\right)}^2\right)+6\hatx \left(2\hatx-1\right) {\hatz}^2 \log(\hatz)\Biggl]\,,\\ \Delta\hat{\sigma}_2^{qq} &= \frac{ \left(N_c^2-1\right)\hatx }{{N_c}^2 q_T \left(1-\hatz\right)} \Biggl[ \left(1-\hatz\right) \bigl(\left(1+N_c^2\right) \left(-1+\hatx\right)+\left(N_c^2-1\right) \left(1-3\hatx\right) \hatz\bigr)+2 \left(2\hatx-1\right) \hatz \log(\hatz)\Biggr]\,, \\ \Delta\hat{\sigma}_3^{qq} & =\frac{\left(N_c^2-1\right) \hatx}{ 4N_c^2 Q \left(1-\hatx\right) {\left(1-\hatz\right)}^2} \Biggl[\left(1-\hatz\right) \Bigl\{\left(1-\hatx\right) \left(5\hatx+N_c^2 \left(2-11\hatx\right)\right) \hatz-\left(1+N_c^2\right) (1-\hatx)^2 \\ & \qquad \qquad-\left(N_c^2-1\right) \left(1+\hatx \left(14\hatx-13\right)\right) {\hatz}^2\Bigr\}-2\hatz \left(1-\hatz-\hatx \left(5-4\hatx-8\left(1-\hatx\right) \hatz\right)\right) \log(\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_4^{qq} & =\frac{\left(N_c^2-1\right) \hatx}{ 2N_c^2 q_T {\left(1-\hatz\right)}^2} \Biggl[\left(1-\hatz\right) \Bigl\{3{\left(1-\hatz\right)}^2+\hatx \left(-3+\left(5-4\hatz\right) \hatz\right) \\ & \qquad \qquad-N_c^2 \left(1-\hatz\right) \left(2-3\hatz+\hatx \left(-2+4\hatz\right)\right)\Bigr\}-2\left(\hatx-{\left(1-\hatz\right)}^2-2\hatx \left(1-\hatz\right) \hatz\right) \log(\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_8^{qq}& =\frac{ (N_c^2-1)\hatx }{ 4N_c^2 Q \left(1-\hatx\right) {\left(1-\hatz\right)}^2} \Biggl[ \left(1+N_c^2\right) {\left(1-\hatx\right)}^2+\left(1-\hatx\right) \left(5-10\hatx+N_c^2 \left(6\hatx-5\right)\right) \hatz\\ & \qquad \qquad -\left(9+\hatx \left(15\hatx-26\right)-N_c^2 \left(7+9\left(\hatx-2\right) \hatx\right)\right) {\hatz}^2 \\ & \qquad \qquad -\left(N_c^2-1\right) \left(3+\hatx \left(4\hatx-9\right)\right) {\hatz}^3-2\hatz \left(\hatx \left(7-4\hatx-2\hatz\right)+\hatz-3\right) \log(\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_9^{qq} & =\frac{\left(N_c^2-1\right) \hatx }{ 2N_c^2 q_T {\left(1-\hatz\right)}^2} \Biggl[ \left(1-\hatz\right) \Bigl(2N_c^2 \left(1-\hatx \left(1-\hatz\right)-2\hatz\right) \left(1-\hatz\right)-3+\hatz \left(9-4\hatz\right)+\hatx \left(3-\left(7-2\hatz\right) \hatz\right)\Bigr) \\ & \qquad \qquad -2\left(1+\left(\hatz-3\right) \hatz+\hatx \left(2\hatz-1\right)\right) \log(\hatz)\Biggr]\,,\\ \end{split}\] and in the gluon-fragmenting (\(qg\)) channel by \[\begin{split} \Delta\hat{\sigma}_{D8}^{qg} & =\frac{ \left(N_c^2-1\right) \hatx }{2N_c^2 Q \hatz} \Biggl[\hatz \left(-2+\hatz-N_c^2 \hatz+\hatx \left(4-3\hatz+N_c^2 \left(-2+3\hatz\right)\right)\right)-2\left(1-2\hatx\right) \left(1-\hatz\right) \log(1-\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_{D9}^{qg} & =\frac{ \left(N_c^2-1\right) \left(1-\hatx\right) \hatx \left(1-\hatz\right) }{2{N_c}^2 q_T {\hatz}^2} \Biggl[ \hatz \left(2+3\left(N_c^2-1\right) \hatz\right)+\left(2-4\hatz\right) \log(1-\hatz)\Biggr]\,, \end{split}\] \[\begin{split} \Delta\hat{\sigma}_1^{qg} & = \frac{N_c^2-1}{2N_c^2 q_T {\hatz}^2} \Biggl[\hatz \left(1+{\hatx}^2 \left(-13+23\hatz-10{\hatz}^2\right)+3\hatx \left(2-3\hatz+{\hatz}^2\right)\right) \\ & \qquad \qquad-N_c^2 \left(1-\hatz\right) \left(-1-3\hatx {\hatz}^2+{\hatx}^2 \left(1-8\hatz+10{\hatz}^2\right)\right)-6\hatx \left(-1+2\hatx\right) {\left(1-\hatz\right)}^2 \log(1-\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_2^{qg} & = \frac{\left(N_c^2-1\right) \hatx \left(1-\hatz\right) }{N_c^2 q_T {\hatz}^2} \Biggr[\hatz \Bigl(2-\hatz+N_c^2 \hatz-\hatx \left(4-3\hatz+{N_c}^2 \left(-2+3\hatz\right)\right)\Bigr)+2\left(1-2\hatx\right) \left(1-\hatz\right) \log(1-\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_3^{qg} & =-\frac{\left(N_c^2-1\right) \hatx}{4{N_c}^2 Q \left(1-\hatx\right) {\hatz}^2} \Biggl[\hatz \Bigl\{\left(-1+\hatz\right) \left(2+\left(N_c^2-1\right) \hatz\right)+\hatx \left(10-24\hatz+13{\hatz}^2+N_c^2 \left(-4+16\hatz-13{\hatz}^2\right)\right) \\ & \qquad \qquad +{\hatx}^2 \left(-8+23\hatz-14{\hatz}^2+N_c^2 \left(4-17\hatz+14{\hatz}^2\right)\right)\Bigr\} \\ & \qquad \qquad +2\left(1-\hatz\right) \left(-1+\hatx \left(5-8\hatz\right)+\hatz+{\hatx}^2 \left(-4+8\hatz\right)\right) \log(1-\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_4^{qg} & =-\frac{\left(N_c^2-1\right) \hatx \left(1-\hatz\right) }{2{N_c}^2 q_T {\hatz}^3} \Biggl[\hatz \left(2+\left(4N_c^2-3\right) \hatz-3\left(N_c^2-1\right) {\hatz}^2+\hatx \left(-2+\left(3-2{N_c}^2\right) \hatz+4\left(N_c^2-1\right) {\hatz}^2\right)\right) \\ & \qquad \qquad+2\left({\left(1-\hatz\right)}^2-\hatx \left(1-2\hatz+2{\hatz}^2\right)\right) \log(1-\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_8^{qg} & = \frac{\left(N_c^2-1\right) \hatx }{4N_c^2 Q \left(1-\hatx\right) {\hatz}^2} \Biggl[\hatz \Bigl\{6-\left(1+{N_c}^2\right) \hatz+3\left(N_c^2-1\right) {\hatz}^2+\hatx \left(-22+8\hatz+9{\hatz}^2+{N_c}^2 \left(4-9{\hatz}^2\right)\right) \\ & \qquad \qquad +{\hatx}^2 \left(16-9\hatz-4{\hatz}^2+N_c^2 \left(-4+3\hatz+4{\hatz}^2\right)\right)\Bigr\}+2\left(1-\hatz\right) \left(3+8{\hatx}^2+\hatz-\hatx \left(11+2\hatz\right)\right) \log(1-\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_9^{qg} & = \frac{\left(N_c^2-1\right) \hatx \left(1-\hatz\right) }{2N_c^2 q_T {\hatz}^3} \Biggl[\hatz \Bigl(\left(4-N_c^2\right) \hatz-4-4\left(N_c^2-1\right) {\hatz}^2+\hatx \left(4+3\left(N_c^2-2\right) \hatz+2\left(N_c^2-1\right) {\hatz}^2\right)\Bigr) \\ & \qquad \qquad +2\left(-2+2\hatx+3\hatz-4\hatx \hatz+{\hatz}^2\right) \log(1-\hatz)\Biggr]\. \end{split}\] Let us briefly comment on the analytic structure of the above results. The hard kernels depend on the virtual quark propagator \(1/(p_1-l_1)^2\approx 1/(-2p_1^+l_1^-)\). We have parametrized the fragmenting parton momentum as \(p_q=p_2-l_1=P_h/z\) and \(l_1=P_h/z\) in the \(q\to q\) and \(q\to g\) channels, respectively. Since \(P_h^-=z_f q^-\) and \(p_2^-=q^-\), we have \(l_1^-=(1-\hat z)q^-\) in the former case, and \(l_1^-=\hat z q^-\) in the latter case. This is why the hard cross sections for the \(q\to q\) and \(q\to g\) channels contain the factor \(1/(1-\hat z)\) and \(1/\hat z\), respectively. When \(z_f\ll 1\), both factors \(1/(1-\hatz)=z/(z-z_f)\) and \(1/\hatz=z/z_f\) become large as \(z\) is varied between \(z_f\) and 1. When \(z_f\to 1\), only the former becomes large around the endpoint \(z\gtrsim z_f\), so the \(q\to q\) channel dominates over the \(q\to g\) channel. This observation will be confirmed in the later numerical analysis. The denominator \(q_T\) hints that higher-order corrections will introduce the large Sudakov logarithms \(\ln^2(Q/q_T)\) at low \(q_T\), whose resummation should be implemented in principle. This is however beyond the scope of this work.
# Computation of the hard part: Gluon initiated channel
The gluon initiated channel is somewhat simpler, since the right diagram in Fig. [\[fig:gluon\]](#fig:gluon){reference-type="ref" reference="fig:gluon"} does not contribute due to Furry's theorem as already pointed out. We thus consider only four diagrams: the left diagram in Fig. [\[fig:gluon\]](#fig:gluon){reference-type="ref" reference="fig:gluon"} and its crossing diagrams with the photon and gluon attachments being interchanged. Considering the quark-fragmenting channel for definiteness, we sum the four diagrams and their complex-conjugates in the form \[\begin{split} S^{(0)\mu\nu \alpha \beta} (k) & =-\frac{i g^4}{2 (N_c^2-1)} (2\pi)\delta\left((k + q-p_q)^2\right)\int \frac{d^4 l_2}{(2\pi)^4} (2\pi) \delta\left(l_2^2\right) (2\pi) \delta\left((p_2-l_2)^2\right)\\ &\times \left[A^{\alpha\mu}_{lj}(k+q-p_q)\bar{M}_{jikl}(k+q-p_q,l_2) A^{\nu\beta}_{ik}(l_2)-A^{\beta\nu}_{lj}(k+q-p_q)\bar{M}_{jikl}(k+q-p_q,l_2) A^{\mu\alpha}_{ik}(l_2)\right]\,, \end{split} \label{eq:Sg2}\] with \[\begin{split} \bar{M}_{jikl}(k+q-p_q,l_2) & =-\frac{N_c^2-1}{4 N_c} \frac{1}{(k + q-p_q-l_2)^2}\left[\slashed{p}_q \gamma^\rho (\slashed{p}_2-\slashed{l}_2)\right]_{ji} \left[\slashed{l}_2\gamma_\rho (\slashed{k} + \slashed{q}-\slashed{p}_q)\right]_{kl}\,, \label{mbar} \end{split}\] \[A^{\alpha\mu}(k+q-p_q) = \gamma^\alpha \frac{\slashed{p}_q-\slashed{q}}{(p_q-q)^2}\gamma^\mu + \gamma^\mu \frac{\slashed{p}_q-\slashed{k}}{(p_q-k)^2} \gamma^\alpha\,,\] \[A^{\nu\beta}(l_2) = \gamma^\nu \frac{\slashed{p}_1-\slashed{l}_2}{(p_1-l_2)^2}\gamma^\beta + \gamma^\beta \frac{\slashed{q}-\slashed{l}_2}{(q-l_2)^2} \gamma^\nu \. \label{twopole}\] Here \(k\) represents the initial gluon momentum, \(p_q\) is the observed quark, with the unobserved antiquark carrying the momentum \(k+q-p_q\) (equal to \(l_1 = p_2-p_q\) in the collinear limit), and \(l_2\) is the loop momentum. The derivative \(\partial/\partial k^\lambda\) in [\[eq:wglue\]](#eq:wglue){reference-type="eqref" reference="eq:wglue"} can be performed along the steps analogous to the case of the quark initiated channel. Defining \[S^{(0) \mu\nu \alpha \beta} (k) = g^4 (2\pi)\delta\left((k+q-p_q)^2\right)\int \frac{d^2 \vec{l}_{2T}}{(2\pi)^3}\frac{dl_2^+}{2 l_2^+} (2\pi) \delta\left((k + q-l_2)^2\right) \widehat{S}^{(0)\mu\nu\alpha\beta}(k)\,,\] where \(\widehat{S}^{(0)\mu\nu\alpha\beta}(k)\) can be read off from ([\[eq:Sg2\]](#eq:Sg2){reference-type="ref" reference="eq:Sg2"}), we find \[\begin{split} & \frac{d^6\Delta \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} = \frac{\alpha_{\rm em}^2 \alpha_S^2 M_N}{16\pi^2 x_B^2 S_{ep}^2 Q^2} \sum_k \mathcal{A}_k \int \frac{d x}{x} \int \frac{dz}{z} (2\pi)\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2 D_f(z)\\ & \times 2i\Bigg\{ \mathcal{G}_{3T}(x) \int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right) \epsilon^{n\alpha\beta S_T} \widehat{S}_{\mu\nu}^{(0)\alpha'\beta'} (p_1)\omega_{\alpha'\alpha} \omega_{\beta' \beta} \tilde{\mathcal{V}}^{\mu\nu}_k\\ & + x^2\frac{\partial \mathcal{G}_{3T}(x)}{\partial x} \frac{l_{1T}^\beta \epsilon^{\alpha P n S_T}-l_{1T}^\alpha \epsilon^{\beta P n S_T}}{p_1 \cdot l_1} \int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right) \widehat{S}_{\mu\nu\alpha\beta}^{(0)}(p_1)\tilde{\mathcal{V}}^{\mu\nu}_k \\ & + x\mathcal{G}_{3T}(x) \frac{\partial}{\partial x}\Bigg\{x\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right)\left[\frac{l_{1T}^\beta \epsilon^{\alpha P n S_T}-l_{1T}^\alpha \epsilon^{\beta P n S_T}}{p_1\cdot l_1} + \frac{l_{2T}^\beta \epsilon^{\alpha P n S_T}-l_{2T}^\alpha \epsilon^{\beta P n S_T}}{p_1\cdot (p_2-l_2)}\right] \widehat{S}_{\mu\nu\alpha\beta}^{(0)}(p_1)\tilde{\mathcal{V}}^{\mu\nu}_k\Bigg\}\\ &-x\mathcal{G}_{3T}(x)\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2l_2^+} (2\pi)\delta\left((p_2-l_2)^2\right) \frac{l_{2T}^\beta \epsilon^{\alpha P n S_T}-l_{2T}^\alpha \epsilon^{\beta P n S_T}}{p_1\cdot (p_2-l_2)} x\frac{\partial}{\partial x} \widehat{S}_{\mu\nu\alpha\beta}^{(0)}(p_1)\tilde{\mathcal{V}}^{\mu\nu}_k\\ &-x\mathcal{G}_{3T}(x)\int \frac{d^2 \vec{l}_{2T} d l_2^+}{(2\pi)^3 2 l_2^+}(2\pi) \delta\left((p_2-l_2)^2\right)\left(g_T^{\beta\lambda} \epsilon^{\alpha P n S_T}-g_T^{\alpha\lambda} \epsilon^{\beta P n S_T}\right)\left(\frac{\partial}{\partial k^\lambda} \widehat{S}_{\mu\nu\alpha\beta}^{(0)}(k)\right)_{k = p_1} \tilde{\mathcal{V}}^{\mu\nu}_k\Bigg\}\. \end{split} \label{eq:pxg2}\] It will be useful to write the 2nd line as \[\epsilon^{n\alpha\beta S_T} \widehat{S}^{(0)\alpha'\beta'}_{\mu\nu} (p_1)\omega_{\alpha'\alpha} \omega_{\beta' \beta} = \epsilon^{nP\beta S_T} \widehat{S}^{(0)}_{\mu\nu n \beta}-\epsilon^{n P \alpha S_T} \widehat{S}^{(0)}_{\mu\nu\alpha n}\.\] Similar to [\[eq:px\]](#eq:px){reference-type="eqref" reference="eq:px"} for the quark-initiated channel, the individual lines in [\[eq:pxg2\]](#eq:pxg2){reference-type="eqref" reference="eq:pxg2"} contain infrared divergences which must be canceled in the sum over all the lines. We will prove this cancellation in Appendix [\[sec:divg1\]](#sec:divg1){reference-type="ref" reference="sec:divg1"}. The hard coefficients can be obtained in complete analogy to the quark initiated channel. We have the same sets of roots \((a_1,a_2)\) as before (see [\[eq:pqpm\]](#eq:pqpm){reference-type="eqref" reference="eq:pqpm"} and [\[eq:l2pm\]](#eq:l2pm){reference-type="eqref" reference="eq:l2pm"}). As we will show in Appendix [\[sec:divg1\]](#sec:divg1){reference-type="ref" reference="sec:divg1"}, a divergence comes neither from the \(p_1-l_2\) propagator (corresponding to the choice \(a_2 =-1\) for the root) nor from the \(q-l_2\) propagator (corresponding to the choice \(a_2 = 1\)) in ([\[twopole\]](#twopole){reference-type="ref" reference="twopole"}). Therefore, the \(l_{2T}\) loop integral is finite, which can be performed analytically. We have also confirmed that, similarly to the previous case, we obtain an expression independent of the choice of the roots for \(l_{1(a_1)}^+\) after the loop integral and after the sum over the \(l_{2(a_2)}^+\) roots. All in all, the final result for the cross section can be written in a compact way as \[\begin{split} & \frac{d^6\Delta \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} = \frac{\alpha_{\rm em}^2 \alpha_S^2 M_N}{16\pi^2 x_B^2 S_{ep}^2 Q^2} \sum_k \mathcal{A}_k \mathcal{S}_k \int^1_{x_{\rm min}} \frac{d x}{x} \int^1_{z_{\rm min}} \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \\ &\qquad \qquad \times\sum_f e_f^2D_f(z) \big[ x^2\frac{\partial \mathcal{G}_{3T}(x)}{\partial x}\Delta\hat{\sigma}^{gq}_{D k} + x \mathcal{G}_{3T}(x) \Delta\hat{\sigma}^{gq}_k\big]\,, \end{split} \label{eq:ddsigg}\] with \(x\partial {\cal G}_{3T}/\partial x \approx-\Delta G(x)/2\) in the present approximation. The hard coefficients are given explicitly by \[\begin{split} \Delta\hat{\sigma}_{D8}^{gq} & = \frac{2\left(1-\hatx\right) \hatx}{N_c Q (1-\hatz)^2 \hatz} \Biggl[(1-\hatx) \Bigl(\hat{z}\log(\hat{z})-(1-\hatz)\log(1-\hatz)\Bigr) + \hatx\hatz(1-\hatz) (1-2\hatz) \Biggr]\,,\\ \Delta\hat{\sigma}_{D9}^{gq} & = \frac{2{\left(1-\hatx\right)}^2 \hatx}{N_c q_T {\left(1-\hatz\right)}^2 {\hatz}^2} \Biggl[{\left(1-\hatz\right)}^2 \log(1-\hatz) + \hatz^2\log(\hatz) +\hatz(1-\hatz)\bigr(\hatz^2+(1-\hatz)^2\bigr) \Biggr]\,, \end{split}\] \[\begin{split} \Delta\hat{\sigma}_1^{gq} & = \frac{\left(1-\hatx\right) }{N_c q_T \left(1-\hatz\right) {\hatz}^2} \Biggl[\left(1-2\hatz\right) \left(1+2{\hatx}^2 {\left(1-2\hatz\right)}^2-\left(1-\hatz\right) \hatz-2\hatx \left(1-\left(1-\hatz\right) \hatz\right)\right) \\ & \qquad \qquad +6\left(1-\hatx\right) \hatx \bigl(\left(1-\hatz\right) \log(1-\hatz)-\hatz \log(\hatz)\bigr)\Biggr]\,,\\ \Delta\hat{\sigma}_2^{gq} & =\frac{4\left(1-\hatx\right) \hatx}{N_c q_T \left(1-\hatz\right) {\hatz}^2} \Biggl[\left(1-\hatz\right) \bigl(\hatx \left(2\hatz-1\right) \hatz+\left(1-\hatx\right) \log(1-\hatz)\bigr)-\left(1-\hatx\right) \hatz \log(\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_3^{gq} & = \frac{\hatx }{N_c Q {\left(1-\hatz\right)}^2 {\hatz}^2} \Biggl[\left(1-\hatz\right) \hatz \Bigl(1-2\left(1-\hatz\right) \hatz+\hatx \left(-6+\left(13-12\hatz\right) \hatz\right)+{\hatx}^2 \left(5-12\left(1-\hatz\right) \hatz\right)\Bigr) \\ &\qquad \qquad +\left(1-\hatx\right) \Bigl(\left(1-\hatz\right) \left(1-\hatz+\hatx \left(-3+4\hatz\right)\right) \log(1-\hatz)-\hatz \left(1-\hatz-\hatx \left(1-4\hatz\right)\right) \log(\hatz)\Bigr)\Biggr]\,,\\ \Delta\hat{\sigma}_4^{gq} & = \frac{2\left(1-\hatx\right) \hatx }{N_c q_T {\left(1-\hatz\right)}^2 {\hatz}^3} \Biggl[\left(1-\hatz\right) \Bigl\{\hatz \Bigl(\hatx-2\hatx \hatz \left(2+\hatz \left(-3+2\hatz\right)\right)-\left(1-\hatz\right) \left(1+\hatz \left(-3+4\hatz\right)\right)\Bigr) \\ & \qquad \qquad -\left(1-\hatx \left(1-\hatz\right)\right) \left(1-\hatz\right) \log(1-\hatz)\Bigr\}-\hatx {\hatz}^3 \log(\hatz)\Biggr]\,,\\ \Delta\hat{\sigma}_8^{gq} & = \frac{\hatx }{N_c Q {\left(1-\hatz\right)}^2 {\hatz}^2} \Biggl[\left(-1+\hatz\right) \hatz \left(-3+10\hatx-7{\hatx}^2+\left(-2+\hatx\right) \hatz+2\left(1+2\left(-1+\hatx\right) \hatx\right) {\hatz}^2\right) \\ & \qquad \qquad +\left(1-\hatx\right) \Bigl(\left(1-\hatz\right) \left(3-3\hatz+\hatx \left(-5+2\hatz\right)\right) \log(1-\hatz)-\hatz \left(1-3\hatz+\hatx \left(-1+2\hatz\right)\right) \log(\hatz)\Bigr)\Biggr]\,,\\ \Delta\hat{\sigma}_9^{gq} & = \frac{-2\left(1-\hatx\right) \hatx }{N_c q_T {\left(1-\hatz\right)}^2 {\hatz}^3} \Biggl[\left(2+\hatx \left(-2+\hatz\right)-2\hatz\right) {\left(1-\hatz\right)}^2 \log(1-\hatz)+\hatz^2 \left(1-\hatx-\left(2-\hatx\right) \hatz\right) \log(\hatz) \\ & \qquad \qquad \qquad \qquad-\hatz \left(1-\hatz\right) \left(-2+3\hatz+\left(-1+\hatz\right) \left(-2\hatx+\hatx \hatz+2\left(-1+\hatx\right) {\hatz}^2\right)\right)\Biggr]\. \end{split}\] In this computation we have chosen the quark to be observed in the final state (\(p_q \to P_h/z\)) while the antiquark goes unobserved. As a nontrivial check we have verified that taking the antiquark as the observed final state (\(l_1 \to P_h/z\)) and the quark as the unobserved one, we recover exactly the same hard coefficients.
# Numerical results
With all the analytical results presented in the previous sections, we are now ready to make predictions for EIC measurements. Specifically, we will numerically compute the asymmetries from the following definition \[A_{UT}^{\sin(\alpha\phi_h + \beta \phi_S)} = \frac{2 \int_0^{2\pi} d\phi_h d\phi_S \sin(\alpha \phi_h + \beta \phi_S) \left[d\sigma(\phi_h,\phi_S)-d\sigma(\phi_h,\phi_S + \pi)\right]}{\int_0^{2\pi} d\phi_h d\phi_S\left[d\sigma(\phi_h,\phi_S)+d\sigma(\phi_h,\phi_S + \pi)\right]}\,, \label{eq:moments}\] where \(d\sigma(\phi_h,\phi_S)\) is a shorthand for \[d\sigma(\phi_h,\phi_S) \equiv \frac{d^6 \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi}\.\] The numerator of [\[eq:moments\]](#eq:moments){reference-type="eqref" reference="eq:moments"} is proportional to the \({\cal O}(\alpha_s^2)\) polarized cross section we calculated. In terms of the Fourier coefficients ([\[fourier\]](#fourier){reference-type="ref" reference="fourier"}), we have \[\begin{split} \mathcal{F}_1 & = \alpha_S M_N \mathcal{F}_0\int \frac{dx}{x}\int \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2 \\ &\times \big[(1+\cosh^2\psi) \left(D_f(z)x g_{Tf}(x)\Delta\hat{\sigma}_1^{qq} + D_f(z)x \mathcal{G}_{3T}(x)\Delta\hat{\sigma}_1^{gq} + D_g(z)x g_{Tf}(x)\Delta\hat{\sigma}_1^{qg}\right)\\ &-2\left(D_f(z)xg_{Tf}(x)\Delta\hat{\sigma}_2^{qq} + D_f(z)x\mathcal{G}_{3T}(x)\Delta\hat{\sigma}_2^{gq} + D_g(z)xg_{Tf}(x)\Delta\hat{\sigma}_2^{qg} \right)\big]\,,\\ \mathcal{F}_2 & = \alpha_S M_N \mathcal{F}_0 \int \frac{dx}{x}\int \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2(-\sinh 2\psi)\\ & \times\left[D_f(z) x g_{Tf}(x)\Delta\hat{\sigma}_3^{qq} + D_f(z) x \mathcal{G}_{3T}(x)\Delta\hat{\sigma}_3^{gq} + D_g(z)x g_{Tf}(x)\Delta\hat{\sigma}_3^{qg}\right]\,,\\ \mathcal{F}_3 & = \alpha_S M_N \mathcal{F}_0 \int \frac{dx}{x}\int \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2 \sinh^2\psi\\ &\times\left[\left(D_f(z) x g_{Tf}(x)\Delta\hat{\sigma}_4^{qq} + D_f(z) x \mathcal{G}_{3T}(x)\Delta\hat{\sigma}_4^{gq} + D_g(z)x g_{Tf}(x)\Delta\hat{\sigma}_4^{qg}\right)\right]\,,\\ \mathcal{F}_4 &= \alpha_S M_N \mathcal{F}_0 \int \frac{dx}{x}\int \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2(-\sinh 2\psi)\\ &\times\Big[D_f(z)x^2 g'_{Tf}(x)\Delta\hat{\sigma}_{D8}^{qq} + D_f(z)x^2 \mathcal{G}'_{3T}(x)(x)\Delta\hat{\sigma}_{D8}^{gq} + D_g(z)x^2 g'_{Tf}(x)\Delta\hat{\sigma}_{D8}^{qg}\\ & + D_f(z)xg_{Tf}(x)\Delta\hat{\sigma}_8^{qq} + D_f(z)x \mathcal{G}_{3T}(x)\Delta\hat{\sigma}_8^{gq} + D_g(z)xg_{Tf}(x)\Delta\hat{\sigma}_8^{qg}\Big]\,,\\ \mathcal{F}_5 & = \alpha_S M_N \mathcal{F}_0 \int \frac{dx}{x}\int \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right) \sum_f e_f^2\sinh^2 \psi\\ &\times\Big[D_f(z)x^2 g'_{Tf}(x)\Delta\hat{\sigma}_{D9}^{qq} + D_f(z)x^2 \mathcal{G}'_{3T}(x)\Delta\hat{\sigma}_{D9}^{gq} + D_g(z)x^2 g'_{Tf}(x)\Delta\hat{\sigma}_{D9}^{qg}\\ & + D_f(z)xg_{Tf}(x)\Delta\hat{\sigma}_9^{qq} + D_f(z)x\mathcal{G}_{3T}(x)\Delta\hat{\sigma}_9^{gq} + D_g(z)xg_{Tf}(x)\Delta\hat{\sigma}_9^{qg}\Big]\,,\\ \end{split}\] with the definition \[\begin{aligned}
{\cal F}_0= \frac{\alpha_{\rm em}^2\alpha_S}{16\pi^2x_B^2S_{ep}^2Q^2}\,, \end{aligned}\] and the abbreviations \[\begin{aligned}
\qquad xg'_{Tf}(x)\equiv x\frac{ \partial g_{Tf}(x)}{\partial x} \approx-\Delta q_f(x)\,, \qquad x{\cal G}'_{3T}(x)\equiv x \frac{ \partial {\cal G}_{3T}(x) }{\partial x}\approx-\frac{\Delta G(x)}{2}\. \end{aligned}\] The integration variables \(x\) and \(z\) are in the ranges \[\begin{aligned}
1>x>x_{\rm min} \equiv x_B\left(1+\frac{z_f}{1-z_f}\frac{q_T^2}{Q^2}\right)\,, \qquad 1>z>z_{\rm min}\equiv z_f\left(1+\frac{x_B}{1-x_B}\frac{q_T^2}{Q^2}\right)\. \label{range} \end{aligned}\] As for the unpolarized cross section in the denominator, we use the leading-order \({\cal O}(\alpha_s)\) formula, summarized by Eqs. (54)-(59) in. The angular decomposition can be cast in the following form \[\frac{d^6 \sigma}{dx_B dQ^2 d z_f d q_T^2 d\phi d \chi} = F^{1} + F^{\cos\phi_h}\cos\phi_h + F^{\cos 2\phi_h} \cos 2\phi_h\,,\\ \] Since we integrate over the lepton angle (see below), only the first term \(F^1\) is relevant with the explicit expression \[\begin{split} F^1 & = {\cal F}_0 \int \frac{dx}{x}\int \frac{dz}{z}\delta\left(\frac{q_T^2}{Q^2}-\left(1-\frac{1}{\hatx}\right)\left(1-\frac{1}{\hatz}\right)\right)\\ &\times \sum_f e_f^2 \Big\{D_f(z)q_f(x) \left[(1+\cosh^2\psi)\hat{\sigma}_1^{qq}-2\hat{\sigma}_2^{qq}\right] + D_g(z)q_f(x) \left[(1+\cosh^2\psi)\hat{\sigma}_1^{qg}-2\hat{\sigma}_2^{qg}\right]\\ & \qquad \qquad + D_f(z)G(x) \left[(1+\cosh^2\psi)\hat{\sigma}_1^{gq}-2\hat{\sigma}_2^{gq}\right]\Big\}\,,\\ \end{split} \label{unpol}\] where \(G(x)\) is the unpolarized gluon PDF and the summation over \(f\) includes both quarks and antiquarks. The hard factors are given by \[\begin{split} & \hat{\sigma}_1^{qq} = \frac{N_c^2-1}{N_c}\hatx\hatz\left\{\frac{1}{Q^2 q_T^2}\left[\frac{Q^4}{\hatx^2 \hatz^2} + (Q^2-q_T^2)^2\right] + 6\right\}\,,\\ & \hat{\sigma}_2^{qq} = 2 \hat{\sigma}_4^{qq} = 4\frac{N_c^2-1}{N_c} \hatx\hatz\,, \end{split}\] \[\begin{split} & \hat{\sigma}_1^{qg} = \frac{N_c^2-1}{N_c}\hatx(1-\hatz)\left\{\frac{1}{Q^2 q_T^2}\left[\frac{Q^4}{\hatx^2 \hatz^2} + \frac{(1-\hatz^2)}{\hatz^2}\left(Q^2-\frac{\hatz^2}{(1-\hatz)^2}q_T^2\right)^2\right] + 6\right\}\,,\\ & \hat{\sigma}_2^{qg} = 2 \hat{\sigma}_4^{qg} = 4\frac{N_c^2-1}{N_c} \hatx(1-\hatz)\,, \end{split}\] \[\begin{split} & \hat{\sigma}_1^{gq} = \hatx(1-\hatx)\left[\frac{Q^2}{q_T^2}\left(\frac{1}{\hatx^2 \hatz^2}-\frac{2}{\hatx \hatz} + 2\right) + 10-\frac{2}{\hatx}-\frac{2}{\hatz}\right] \,,\\ & \hat{\sigma}_2^{gq} = 2 \hat{\sigma}_4^{gq} = 8 \hatx(1-\hatx) \. \end{split}\] Using \(F^1\) ([\[unpol\]](#unpol){reference-type="ref" reference="unpol"}) and the relation ([\[FF\]](#FF){reference-type="ref" reference="FF"}), we obtain from [\[eq:moments\]](#eq:moments){reference-type="eqref" reference="eq:moments"} \[A_{UT}^{\sin(\alpha \phi_h + \beta \phi_S)} = \frac{F^{\sin(\alpha\phi_h + \beta\phi_S)}}{F^1}\. \label{eq:Auts}\] In practice, we show our results as functions of \(P_{hT}\), \(z_f\) or \(x_B\) by integrating out all the other variables in the numerator and denominator. Instead of \(Q^2\), it is convenient to use \(y = Q^2/(x_B S_{ep})\) so that we have \(dQ^2 = x_B S_{ep} dy\) and the relations \[\begin{split} & 1 + \cosh^2\psi = 2 \frac{1 + (1-y)^2}{y^2}\,,\\ & \sinh^2\psi = 4\frac{1-y}{y^2}\,,\\ & \sinh 2 \psi = 4\frac{(2-y)\sqrt{1-y}}{y^2}\. \end{split}\] There are general kinematical constraints on the integration ranges of \(x_B\), \(z_f\) and \(y\). The condition \(x_{\rm min}<1\) (see ([\[range\]](#range){reference-type="ref" reference="range"})) leads to \[x_B \leq 1-\frac{1}{z_f(1-z_f)}\frac{P_{hT}^2}{y S_{ep}}\. \label{eq:xBmax}\] Requiring the upper bound of \(x_B\) to be positive, we find a condition on \(y\), \[y>\frac{1}{z_f(1-z_f)}\frac{P_{hT}^2}{S_{ep}}\.\] Similar constraints can be obtained from \(z_{\rm min}<1\), which are however not very restrictive because \(P_{hT}^2\ll S_{ep}\). In actual experiments, \(A_{UT}\) is integrated over conveniently chosen bins in \(x_B\), \(z_f\) and \(y\), and we will follow suit. Since we are using the leading-order cross sections for both the numerator and denominator, one may ask a legitimate question about the effect of higher order corrections, in particular when \(P_{hT}\ll Q\) and the resummation of the Sudakov logarithms is required. While such a procedure is well established for unpolarized cross sections, that for transversely polarized cross sections is poorly understood. On a general ground, we expect that the impact of resummation largely cancels in the ratio, but this has to be checked, and will be left for a future work. As for the scale \(\mu\) of the QCD coupling constant \(\alpha_S\) (and also of PDFs and FFs), we argue that the lower scale \(\mu=P_{hT}\) is more appropriate than the larger one \(\mu=Q\) in the typical kinematic region \(P_{hT}\ll Q\) we are considering. This is understood simply from the aspect of the Sudakov (\(k_T\)) resummation usually done in the Fourier conjugate impact parameter space \(b_T\). The running of the coupling tends to pick up a dominant contribution from the large \(b_T\) region under the inverse Fourier transformation (for which some prescription is needed to avoid the Landau pole ). Therefore, the choice of a lower scale \(\mu=P_{hT}\sim O(1/b_T)\) fits the above all-order picture better. The computation is performed with the most recent NNPDF and JAM global fits. For the NNPDF sets we use the helicity PDFs from and FFs from. For the JAM sets we use the helicity PDFs and FFs from a simulatenous fit in. \(g_{Tf}(x)\) and \(\mathcal{G}_{3T}(x)\) are deduced from the helicity PDFs according to the formulas [\[eq:gTww\]](#eq:gTww){reference-type="eqref" reference="eq:gTww"} and [\[eq:G3Tww\]](#eq:G3Tww){reference-type="eqref" reference="eq:G3Tww"} in the WW approximation. The uncertainties in PDFs (FFs) in NNPDF and JAM fits are quantified by the Monte Carlo replica method to generate a variance according to the normal distribution. In all the plots below the band represents a combination of 1-\(\sigma\) uncertainty due to the replica method and also uncertainty in the scale choice according to \(0.5<\mu/P_{hT}< 2.0\), added in quadrature. Note that the edge \(\mu=0.5P_{hT}\) starts at \(P_{hT}=2\) GeV.
## Calculation for COMPASS kinematics
Though our approach is most naturally and legitimately applied to the kinematics for EIC, let us first present the results relevant to the COMPASS experiment . Admittedly, the collinear factorization may not be applicable to the COMPASS kinematics since most of the data points have \(P_{hT}\) below 1 GeV. There is, however, one published data point with \(P_{hT}\approx 1.5\) GeV. We thus only show the \(P_{hT}\) distribution for \(P_{hT}>1\) GeV, integrating out the other variables over the following coverage \[0.003 \leq x_B \leq 0.7\,, \qquad 0.1 \leq y \leq 0.9 \,, \qquad 0.2 \leq z_f \leq 1\,,\] as well as \[Q^2 > 1 \,\, {\rm GeV}^2,\qquad W^2 > 25 \,\, {\rm GeV}^2\,,\] where \(W^2 = (q + P)^2 = Q^2(1-x_B)/x_B\). With a 160 GeV muon beam colliding on a fixed proton target, the center of mass energy is \(\sqrt{S_{\mu p}}\approx 17.4\) GeV. The \(P_{hT}\) distributions are shown in Fig. [\[fig:compass\]](#fig:compass){reference-type="ref" reference="fig:compass"} for both \(\pi^+\) and \(\pi^-\) productions. We see that the Sivers asymmetry for \(\pi^+\) is smaller than \(0.5\%\) in magnitude using the NNPDF fits and about \(\sim 0.5\%-1\%\) in magnitude using the JAM fits. In either case, the sign is opposite to the highest \(P_{hT}\) COMPASS data point (see the top-right plot in Fig. 9 of ). Although the significant experimental uncertainty makes a meaningful comparison difficult, the result does indicate the importance of other sources of SSA, such as the ETQS function. However, \(P_{hT}\gtrsim 1\) GeV is the borderline between the collinear and TMD approaches. Therefore, our analysis implies that not only the Sivers function but also the new higher-twist contributions found in need to be included in the global determination of nonperturbative inputs in this regime. As for the Collins asymmetry, our result is negligibly small. The data show nonvanishing central values at \(P_{hT}=1.5\) GeV (see the top-right plot in Fig. 6 of ), but they are consistent with zero after the large error bars are taken into account.
## Calculation for EIC kinematics
We now present our results for the EIC kinematics. Figure [\[eiczf\]](#eiczf){reference-type="ref" reference="eiczf"} shows the \(z_f\) distribution of the \(\pi^+\) Sivers asymmetry for \(\sqrt{S_{ep}}=45\) GeV integrated over the window \(0.1<x_B<0.9\) and \(0.01<y<0.95\) and \(P_{hT} > 1 \, {\rm GeV}\). The upper bound for the integral over \(P_{hT}\) is obtained from [\[eq:xBmax\]](#eq:xBmax){reference-type="eqref" reference="eq:xBmax"} by placing the remaining kinematic variables at their extremal values in the above kinematic window. We also impose the conditions \(Q^2>1\) GeV\(^2\) and \(W^2=(P+q)^2>25\) GeV\(^2\). In addition to the total asymmetry ('tot'), respective contributions from different channels (\(qq,qg,gq\)) are shown. The asymmetry is largest in the forward region \(z_f\to 1\), at most 1.5% in magnitude, and decreases towards zero as \(z_f\) decreases. As we discussed at the end of Section IV, the large \(z_f\) region is dominated by the quark-fragmenting channel, while the gluon-fragmenting channel becomes important at small \(z_f\). Since the final state quark and gluon are back-to-back, this explains the sign change for the two channels. A somewhat larger asymmetry is observed from the JAM fit than from the NNPDF fit. This is in fact a general feature seen also for example in Fig. [\[fig:compass\]](#fig:compass){reference-type="ref" reference="fig:compass"}, but most directly understood from the \(z_f\)-distributions in Fig. [\[eiczf\]](#eiczf){reference-type="ref" reference="eiczf"} where the \(qg\) channel contribution is dying off more rapidly as \(z_f \to 1\) for the JAM fits. Consequently, the cancellation between the \(qq\) and the \(qg\) channels is less effective using the JAM fits. The underlying reason is the smaller \(g\to \pi^+\) FF in the JAM fit than in the NNPDF fit. Figure [\[eicpt\]](#eicpt){reference-type="ref" reference="eicpt"} gives the \(P_{hT}\) distributions of the \(\pi^+\) Sivers asymmetry in low \(z_f\) (\(0.05 < z_f < 0.4\), left) and high \(z_f\) (\(0.5 < z_f < 0.9\), right) bins. We can again see the role of the \(g\to \pi^+\) FF: at low \(z_f\) the Sivers asymmetry can even become positive (albeit rather small in magnitude) using the NNPDF fits, while in the large \(z_f\) region the \(qg\) channel quickly dies off so the JAM fits predict a larger (negative) Sivers asymmetry, around \(0.5\%\sim 1.5\%\) in magnitude. Note that the gluon-initiated (\(gq\)) channel is negligibly small, almost invisible in the plots. A closer look reveals that the contribution to \(A_N\) from this channel is less than \(10^{-3}\) in all the bins we have studied. We have anticipated that the gluon-initiated channel gives a small contribution to light-hadron production. However, the suppression is stronger than expected, and we attempt to explain the reason in the concluding section. Further predictions for the \(P_{hT}\) distributions of the \(\pi^+\) Sivers asymmetry across three bins in \(x_B\) and \(Q^2\), using the NNPDF and JAM fits, are exhibited in Fig. [\[fig:siverspanelpt\]](#fig:siverspanelpt){reference-type="ref" reference="fig:siverspanelpt"}. We find that the Sivers asymmetry can reach up to \(2\%\) in magnitude for the JAM fit covering both large \(x_B\) (\(0.1 < x_B < 0.7\)) and moderate \(x_B\) (\(0.05 < x_B < 0.1\)) bins for the lowest \(Q^2\) bin. Going from moderate to small \(x_B\), the Sivers asymmetry drops to a sub-percent level, as seen in the last \(x_B\) bin with \(0.001 < x_B < 0.05\). This suppression at small \(x_B\) in fact has the same origin as the smallness of the gluon-initiated channel mentioned above (see the discussion in the concluding section). Figure [\[fig:energypanelpt\]](#fig:energypanelpt){reference-type="ref" reference="fig:energypanelpt"} covers the Sivers asymmetry for two collision energies (\(S_{ep}\)) and two bins in \(x_B\). The results show very mild dropping of the Sivers asymmetry as \(\sqrt{S_{ep}}\) is increased from \(45\) GeV to the top EIC energy of \(140\) GeV (see also ). The reason is that the energy dependence mainly comes from the \(y\) dependence, which roughly cancels out between the numerator and denominator. Compared with an earlier prediction for EIC in the TMD framework at low momentum \(P_{hT}<1\) GeV (see Fig. 21 of ), our result for the Sivers asymmetry is similar or somewhat smaller in magnitude but opposite in sign, although a detailed comparison is not possible because there is no overlap in the plotted \(P_{hT}\) ranges. This suggests that there are cancellations between different mechanisms which may lead to a sign change. However, we emphasize again that when \(P_{hT}\lesssim 1\) GeV, other sources of asymmetries from various twist-three TMDs found in should be added to the contribution from the Sivers function. Finally, in Fig. [\[fig:momentspanelpt\]](#fig:momentspanelpt){reference-type="ref" reference="fig:momentspanelpt"}, we present a full set of moments introduced in [\[FF\]](#FF){reference-type="eqref" reference="FF"} and [\[eq:Auts\]](#eq:Auts){reference-type="eqref" reference="eq:Auts"} for three different bins in \(x_B\) and for fixed bins in \(Q^2\) and \(z_f\) (\(1 < Q^2 < 10 \, \mathrm{GeV}^2\), \(0.5<z_f<0.9\)). We find that two additional moments \(A_{UT}^{\sin(\phi_S)}\) and \(A_{UT}^{\sin(2\phi_h-\phi_S)}\) reach up to \(2\%\) in magnitude in the highest \(x_B\) bin \(0.7 < x_B < 0.1\). In the TMD framework for low\(P_{hT}\), the \(\sin(\phi_S)\) and \(\sin(2\phi_h-\phi_S)\) asymmetries are known to be generated by various twist-three TMDs. We have just demonstrated that the \(g_T\) distribution (or its TMD counterpart \(g_{1T}\) by extension) is also a potentially significant source of these asymmetries. Indeed, our prediction 1-2 % at \(P_{hT}=1\) GeV is comparable to previous TMD-based calculations.
# Discussions and Conclusions
In this paper we have performed the complete analytical and numerical evaluation of the novel two-loop contributions to SSA proportional to the twist-two polarized PDFs \(\Delta q(x)\) and \(\Delta G(x)\). Our results indicate that, at the EIC, \(A_{UT}\) for pions can reach 1-2% for the three harmonics \(\sin (\phi_h-\phi_S)\), \(\sin (\phi_h)\) and \(\sin (2\phi_h-\phi_S)\). On the other hand, contributions from the gluon-initiated channel are negligibly small. Since we are dealing with higher-order perturbative diagrams, we have anticipated that the resulting asymmetry would be small. However, the stronger-than-expected suppression we observed, especially in the gluon sector, calls for an explanation. Parametrically, the asymmetry behaves as \[\begin{aligned}
A_{UT}=\frac{d\Delta \sigma}{d\sigma} \sim \frac{\alpha_s^2 \frac{M_N}{P_{hT}} (x\Delta q(x)\ {\rm or}\ x\Delta G(x))}{\alpha_s (q(x) \ {\rm or} \ G(x))}\. \label{our} \end{aligned}\] In addition to the obvious factor of \(\alpha_s\), \(A_{UT}\) is suppressed by the smallness of polarized PDFs as compared to unpolarized ones. In particular, the gluon-initiated channel is expected to be important for \(x\ll 1\), but there, \(\Delta G(x)\sim xG(x)\) as a rule of thumb. On top of this, there is a somewhat unexpected extra factor of \(x\) in the numerator which comes from the rewriting \(\int dx g_T(x)=\int \frac{dx}{x}xg_T(x)\) and \(\int dx {\cal G}_{3T}(x)=\int \frac{dx}{x}x{\cal G}_{3T}(x)\). Of course the same factor exists in the unpolarized cross section in the denominator, which is, however, accompanied by \(P^+\), and the product \(xP^+=p^+_1\) goes into the hard part and gets absorbed. Therefore, our new contribution, especially in the gluon initiated channel, is strongly suppressed like \(\sim x^2\) at low \(x\), or in more practical terms, as the selected kinematic bin is sensitive to the low \(x_B\) region. This tendency has been clearly shown in Fig. [\[fig:momentspanelpt\]](#fig:momentspanelpt){reference-type="ref" reference="fig:momentspanelpt"}. In the literature, gluon-initiated channels are usually ignored in the calculation of \(A_{UT}\) for light-hadrons (see, however, an attempt in \(pp\) collisions ), partly because it is believed to be small, but also because nothing is known about the strength of the three-gluon correlators \(\langle FgFF\rangle\). For the first time, we have presented a reliably calculable piece of the gluon initiated contributions, and found very small values. After all, our main interest in the gluon initiated processes focuses on \(A_{UT}\) in heavy systems such as open charm and \(J/\psi\). This will be studied elsewhere. It is worthwhile to compare ([\[our\]](#our){reference-type="ref" reference="our"}) with the well-known parametric estimate of SSA \[\begin{aligned}
A_{UT}\sim \frac{\alpha_s m_q}{P_{hT}}\,, \label{kpr} \end{aligned}\] where \(m_q\sim\) a few MeV is the current quark mass. This formula has been inferred from the argument in, and is often quoted in order to emphasize the smallness of SSA in perturbation theory and the necessity to introduce new nonperturbative distributions. The factor of \(\alpha_s\) is because one needs loop diagrams such as in Fig. 1 to get an imaginary part, and the factor of \(m_q\) is because one needs a helicity flip. However, this suppression by \(m_q\) is illusory for the proton initial state. As is clear from the definitions of \(g_T\) and \({\cal G}_{3T}\) in ([\[gtdef\]](#gtdef){reference-type="ref" reference="gtdef"}) and ([\[eq:G3T\]](#eq:G3T){reference-type="ref" reference="eq:G3T"}), \(m_q\) is replaced by the proton mass \(M_N\) (see also a related argument in ). Thus the correct argument in the DIS case would be that, naively \(A_{UT}\sim \frac{\alpha_s M_N}{P_{hT}}\) is large, but the coefficient is suppressed due to the above-mentioned factor \(x^2\), resulting in SSA of about 1% as we have shown. In SIDIS at \(P_{hT}>1\) GeV, this should be comparable to other nonperturbative origins of SSA. Precisely measuring \(A_{UT}\) in the sub-percent region is challenging at the EIC. Conversely, if the future data on \(A_{UT}\) turn out to be consistently larger than 1%, most likely genuine twist-three effects are at work. But our result must be subtracted when one tries to extract various twist-three distributions. The distinct kinematical features of our contribution, such as the suppression in low \(z_f\) and low-\(x_B\) regions, may be useful to isolate this purely perturbative 'background'. At lower \(P_{hT}<1\) GeV, predictions based on the Sivers function are available . However, in the TMD regime \(P_{hT}<1\) GeV, there are many other sources of the \(\sin (\phi_h-\phi_S)\) asymmetry which are unrelated to the Sivers function, that must be taken into consideration in order to reliably extract the Sivers function. We thank Shinsuke Yoshida for useful discussions. S. B. would like to thank for the warm hospitality of the Yukawa Institute for Theoretical Physics, Kyoto University where part of this work was performed for which S. B. was supported by the JSPS postdoctoral fellowship for foreign researchers under Grant No. 17F117323. S. B. and A. K. are supported by the Croatian Science Foundation (HRZZ) no. 5332 (UIP-2019-04). The work by Y. H. is supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, under contract number DE-SC0012704, and also by Laboratory Directed Research and Development (LDRD) funds from Brookhaven Science Associates. H. n. L is supported by the Ministry of Science and Technology of R.O.C. under Grant No. MOST-110-2811-M-001-540-MY3.
[^1]: \(\Delta \bar{q}(x)\) is formally related to the operator definition of \(g_T(x)\) in the negative support region \(0>x>-1\). We have checked that the antiquark contribution can be effectively included via the replacement ([\[eq:gTww\]](#eq:gTww){reference-type="ref" reference="eq:gTww"}) in the physical region \(1>x>0\) using the same hard kernel.
[^2]: To avoid confusion, we note that the \(G_F\) and \(\tilde{G}_F\) pieces in ([\[eq:wtot\]](#eq:wtot){reference-type="ref" reference="eq:wtot"}) are not the conventional ETQS contribution quoted here, but rather its \({\cal O}(\alpha_s)\) corrections. | {'timestamp': '2021-09-14T02:19:42', 'yymm': '2109', 'arxiv_id': '2109.05440', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05440'} |
# Discussion and Conclusion {#sec:conclusion}
In this paper, we propose *Confirm-it*, a multimodal conversational model based on a decoding strategy inspired by cognitive studies of human behavior. We show that, through the proposed beam search re-reanking algorithm, our model generates dialogues that are more effective (based on task-accuracy) and more natural (based on the dialogues features discussed above). We believe further improvement could be obtained by increasing the performance of every single module. Moreover, the structure of the generated dialogues remains to be analysed, and we agree with that a proper evaluation should involve humans. In future work, our method can be easily extended to other task-oriented dialogue tasks which involve a conversational agent as far as it has a module that generates questions and a module that performs a classification task. Depending on the task at hand, different ways to take intermediate probabilities into account can be designed, but the core idea of the method would not change.
# Introduction {#sec:introduction}
Quite important progress has been made on multimodal conversational systems thanks to the introduction of the Encoder-Decoder framework . The success of these systems can be measured by evaluating them on task-oriented referential games. Despite the high task-success achieved and the apparent linguistic well-formedness of the single questions, the quality of the generated dialogues, according to surface-level features, have been shown to be poor; this holds for systems based on both greedy and beam search (e.g. ). found that when taking these surface-level features as a proxy of linguistic quality, the latter does not correlate with task success, and the authors point to the importance of studying deeper features of dialogue structures. We aim to develop a multimodal model able to generate dialogues that resemble human dialogue strategies. Cognitive studies show that humans do not always act as "rational" agents. When referring to objects, they tend to be overspecific and prefer properties irrespectively of their utility for identifying the referent ; when searching for information or when learning a language, they tend to follow confirmation-driven strategies. Modelling such behaviour in language learning, and propose a procedure in which a single hypothesized word-referent pair is maintained across learning instances, and it is abandoned only if the subsequent instance fails to confirm the pairing. Inspired by these theories, we propose a model, *Confirm-it*, which generates questions driven by the agent's confirmation bias. Take the example of a referential guessing game in which an agent has to ask questions to guess an object in a given image. *Confirm-it* will ask questions that reinforce its beliefs about which is the target object, till proven otherwise. For instance, in Figure [\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}, after learning that the target is a living entity (turn 1), the agent conjectures the target is the dog on the right of the picture (though in principle, it could have been any of the candidates). Hence, the decoder generates the question that would let it confirm such belief, "is it a dog?". If its expectations are not met (viz., it receives a negative answer to such question-turn 2b), it moves its attention to another candidate object. We do not claim that our choice represents the optimal strategy to play the game, but we believe that it makes the generated dialogue more human-like. To evaluate this strategy, we take as a test-bed GuessWhat?!, a two-player game between a Questioner that has to guess the target, and an Oracle (called "external Oracle" in the following) who is aware of the target. The widely used architecture of the Questioner, GDSE, jointly trains a Question Generator (QGen) and a Guesser . We augment this architecture with a module that simulates an internal Oracle. Being an "internal" Oracle, at test time this agent does not know what the target object is: while at training time it learns to answer questions by receiving the gold standard datapoint (the question, the actual target, and the human answer), at test time it assumes the target is the candidate object to which the Guesser assigns the highest probability. Hence, the three modules of the Questioner straightforwardly cooperate one another. The internal Oracle guides the QGen to ask questions that reinforce the Guesser's beliefs. Concretely, at training time, through Supervised Learning (SL) the QGen learns to ask human-like questions turn-by-turn, the internal Oracle to answer them, and the Guesser to guess the target object once the dialogue ends. At test time, we implement a beam search re-ranking algorithm that simulates the single-conjecture learning strategy used by humans: among the questions the QGen generates via beam search, the algorithm promotes the questions whose answer (obtained via the internal Oracle that receives the candidate with the highest probability as the target) increases the most the model's confidence in its hypothesis about the target. We run both quantitative and qualitative analyses, and evaluate the effectiveness of the dialogue strategy by asking human annotators to guess the target object given the dialogues generated by *Confirm-it*. We compare results giving the dialogue generated by *Confirm-it* when using the re-ranking algorithm and when generating the question proposed by the plain beam search. We show that the task accuracy of both the conversational agent and human subjects increases when receiving the dialogues generated by the *Confirm-it* re-ranking algorithm.
# Introduction {#sec:introduction}
Quite important progress has been made on multimodal conversational systems thanks to the introduction of the Encoder-Decoder framework . The success of these systems can be measured by evaluating them on task-oriented referential games. Despite the high task-success reached and the apparent linguistic well-formedness of the single questions, the quality of the generated dialogues, according to surface-level features, have been shown to be poor; this holds for systems based on both greedy and beam search . found that when taking these surface-level features as a proxy of linguistic quality, the latter does not correlate with task success, and the authors point to the importance of studying deeper features of dialogue structures. We aim to develop a multimodal model able to generate more natural dialogues, namely dialogues that resemble human dialogue strategies. Cognitive studies show that humans do not always act as "rational" agents. When referring to objects, they tend to be overspecific and prefer properties irrespectively of their utility for identifying the referent ; when searching for information or when learning a language, they tend to follow confirmation-driven strategies. Modelling such behaviour in language learning, and propose a procedure in which a single hypothesized word-referent pair is maintained across learning instances, and it is abandoned only if the subsequent instance fails to confirm the pairing, reflecting a 'fast mapping' learning process. Inspired by these theories, we propose a model, *Confirm-it*, which generates questions driven by the agent's confirmation bias. Take the example of a referential guessing game in which an agent has to ask questions to guess an object in a given image. *Confirm-it* will ask questions that reinforce its beliefs about which is the target object, till proven otherwise. For instance, in Figure [\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}, after learning that the target is a living entity (turn 1), the agent conjectures the target is the dog on the right of the picture (though in principle, it could have been any of the candidates). Hence, the decoder generates the question that would let it confirm such belief, "is it a dog?". If its expectations are not met (viz., it receives a negative answer to such question-turn 2b), it moves its attention to another candidate object. We do not claim that our choice represents the optimal strategy to play the game, but we believe that it makes the dialogue more natural in its flow and more effective because it takes into account the model's intrinsic biases and intermediate predictions. To evaluate this strategy, we take as a test-bed GuessWhat?!, a two-player game between a Questioner that has to guess the target and an Oracle. The widely used architecture of the Questioner, GDSE, jointly train a Question Generator (QGen) and a Guesser . We augment it with a module that simulates an internal oracle. The three modules of the Questioner straightly cooperate one with the other: the internal oracle guides the question generator to ask questions that reinforce the Guesser's beliefs. Concretely, at training time, through Supervised Learning (SL) the QGen learns to ask human-like questions turn-by-turn, the internal oracle to answer them, and the guesser to guess the target object once the dialogue ends. At test time, we exploit a beam search re-ranking algorithm that simulates the single-conjecture learning strategy used by humans. Among the questions the QGen generates via beam search, the algorithm promotes the one that helps the guesser reinforce its belief. We run both quantitative and qualitative analyses, and evaluate the effectiveness of the dialogue strategy by asking humans to guess the target given dialogues generated by *Confirm-it*. We compare results when giving the dialogue generated by *Confirm-it* using the re-ranking algorithm and when generating the question proposed by the plain beam search. We show that the task accuracy both of the conversational agent and human subjects increases when receiving dialogues generated by the *Confirm-it* re-ranking.
# Model and Re-ranking Strategy {#sec:model_experiments}
Our model, *Confirm-it*, builds on GDSE . In the latter, the hidden state representation produced by a multimodal encoder is used to jointly train the question generator (QGen) and the Guesser module. The image is encoded with a ResNet-152 network and the dialogue history is encoded via an LSTM network. QGen uses greedy search to generate questions. To this multi-tasking setting, *Confirm-it* adds an internal Oracle trained to answer the question at each turn. Moreover, it relies on beam search and, at inference time, it goes through a re-ranking phase which simulates the single-conjecture learning strategy. The model architecture is provided in the Supplementary Material (SM) and the algorithm is spelled out below.
We study to what extent the re-ranking phase lets the model generate more effective and more natural dialogues. To this end, we evaluate the *Confirm-it* task-accuracy with and without the ranking phase[^2] and report qualitative analyses of the dialogues.
#### Task-accuracy
Table [1](#tab:accuracy){reference-type="ref" reference="tab:accuracy"} shows the task accuracy results of our model in the GuessWhat?!game. When the model undergoes the re-ranking phase, *Confirm-it* accuracy has an increase of +4.35% with respect to what it achieves when it outputs the question selected by the plain beam search, and an increase of +4.8% against greedy search. Note that, instead, randomly re-ranking the set of questions lowers the performance. This result shows that confirmation-driven strategies help generate more effective dialogues. Interestingly, our re-ranking method does not require additional training compared to the SL paradigm.
#### More Effective Dialogues
To verify whether the improvement of *Confirm-it* is really due to the generation of more effective dialogues to solve the guessing task, we asked human subjects to guess the target given a dialogue. We sampled 500 games from the GuessWhat?!test set containing less than 7 candidate objects. Each participant played 150 games equally divided among dialogues generated by the model with the plain beam search, with our re-ranking strategy, and by the GuessWhat?!human players (taken from the original test set). We made sure no participant played the same game more than once. In total, 10 English proficient volunteers within our organization joined the experiment. As we can see from Table [\[tab:analysis\]](#tab:analysis){reference-type="ref" reference="tab:analysis"}, human annotators reach an accuracy of 70.8% in identifying the target object when receiving dialogues generated by beam search and 77% with *Confirm-it*, suggesting that the re-ranking phase let the model generate more effective dialogues. The accuracy that the annotators achieve when playing the game with dialogues extracted from the original GuessWhat?!test set (and thus generated by human players) is much higher (96%).
#### More Natural Dialogues
Figure [\[fig:comparison\]](#fig:comparison){reference-type="ref" reference="fig:comparison"} reports a sample game that illustrates the difference between a dialogue generated by human players, one generated by the plain beam search, and one by our re-ranking algorithm. The dialogue generated by beam search contains a repetition ("is it a glass?"), it asks about entities not present in the image ("chair" and "glasses") and it ends with a non-conclusive negatively answered question. These features contribute to making the dialogues sound unnatural. We check whether the re-ranking phase helps our model to get closer to human dialogues with respect to these features. To this end, we compute the percentage of games with repeated questions and with the last turn containing a positively answered question. Moreover, we employ CHAIR-s to measure the percentage of hallucinated entities in a sequence, originally proposed in for image captioning and recently applied also to GuessWhat?!. CHAIR-S is defined as the number of dialogues with at last one hallucinated entity divided by the total number of dialogues. As we can see from Table [\[tab:analysis\]](#tab:analysis){reference-type="ref" reference="tab:analysis"}, dialogues generated by *Confirm-it* contain fewer games with at least one repeated question compared to the beam search setting (-8.14%), fewer games with hallucinated entities (-2.61%)[^3], and more games with the last turn containing a positively answered question (71.87% vs. 76.68%). The reduced number of hallucinations is a direct consequence of the *Confirm-it* strategy: following up on a single object through the dialogue, the model is less likely to engage in spurious exchanges on irrelevant objects. Though this strategy continuously looks for confirmations, it is worth noting that it does not increase the number of repetitions, which instead are significantly reduced. This is an interesting property emerging from the interplay between the internal Oracle and the re-ranking strategy, which suggests that asking the very same question more than once in a dialogue does not increase the model's confidence in its hypothesis.
#### Qualitative Analysis of the Strategy
We also evaluated the strategy followed by *Confirm-it* by looking at the model's decisions throughout the dialogue. Interestingly, the model does not select only questions for which it expects a positive answer, though they are the majority (67%). See the SM, for a game in which the re-ranking promoted a question answered negatively by the internal Oracle. Moreover, though the model looks for confirmations, it properly updates its beliefs when disconfirmed: when the model receives from its interlocutor an answer different from the one it expects (based on its internal Oracle), in 70% of the cases the Guesser changes the probabilities over the candidates accordingly, i.e., it assigns the highest probability to a new candidate object. Finally, the use of a human-like strategy does not imply having learned to simply mimic human dialogues from the training set: the re-ranking shows an absolute increase of +12% in the number of pairs of consecutive questions not seen during training (see Table [\[tab:analysis\]](#tab:analysis){reference-type="ref" reference="tab:analysis"}).
# Task and Dataset {#sec:task_data}
GuessWhat?!is an asymmetric game involving two human participants who see a real-world image from MSCOCO. One of the participants (the Oracle) is secretly assigned a target object in the image, while the other participant (the Questioner) has to guess it by asking binary (Yes/No) questions to the Oracle. The GuessWhat?!dataset consists of more than 150k human-human English dialogues containing on average 5.3 questions per dialogue.
[^1]: Our code is available at: <https://github.com/albertotestoni/confirm_it>
[^2]: Remember that *Confirm-it* is based on the GDSE architecture. For comparison, the accuracy reached by GDSE is 45.55% with greedy search and 46.40% using beam search decoding.
[^3]: propose another variant of the CHAIR metric called CHAIR-i (per-instance), defined as the number of hallucinated objects in a sequence divided by the total number of objects mentioned. CHAIR-i results: 18.28 (Beam Search), 15.02 (*Confirm-it*), 4.11 (Human Dialogues). | {'timestamp': '2021-09-14T02:14:22', 'yymm': '2109', 'arxiv_id': '2109.05312', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05312'} |
# Introduction {#sec:introduction}
Learning high-quality sentence embeddings is a fundamental task in Natural Language Processing. The goal is to map semantically similar sentences close together and dissimilar sentences farther apart in the representation space. Many recent successes have been achieved by training on the NLI datasets, where the task is often to classify each sentence pair into one of three categories: entailment, contradiction, or neutral. Despite promising results, prior work share a common weakness: the sentences forming a contradiction pair may not necessarily belong to different semantic categories. Consequently, optimizing the model for semantic entailment and contradiction understanding alone is inadequate to encode the high-level categorical concepts into the representations. Moreover, the vanilla siamese (triplet) loss only learns from the individual sentence pairs (triplets), which often requires substantial training examples to achieve competitive performance. As shown in Section [4.1](#subsec:clustering){reference-type="ref" reference="subsec:clustering"}, the siamese loss can sometimes drive a model to bad local optima where the performance of high-level semantic concept encoding is degraded when compared with its baseline counterpart. In this paper, we take inspiration from self-supervised contrastive learning and propose jointly optimizing the pairwise semantic reasoning objective with an *instance discrimination loss*. We name our approach Pairwise Supervised Contrastive Learning (PairSupCon). PairSupCon leverages this implicit grouping effect to bring together representations from the same semantic category while, simultaneously enhancing the semantic entailment and contradiction reasoning capability of the model. Although the prior work mainly focuses on pairwise semantic similarity related evaluations, we argue in this paper that the capability of encoding the high-level categorical semantic concept into the representations is an equally important aspect for evaluations. As shown in Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"}, the previous state-of-the-art model that performs best on the semantic textual similarity (STS) tasks results in degenerated embeddings of the categorical semantic structure. On the other hand, better capturing the high-level semantic concepts can in turn promote better performance of the low-level semantic entailment and contradiction reasoning. This assumption is consistent with how human categorize objects in a top-down reasoning manner. We further validate our assumption in Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"}, where PairSupCon achieves an averaged improvement of \(10\%-13\%\) over the prior work when evaluated on eight short text clustering tasks, and yields \(5\%-6\%\) averaged improvement on seven STS tasks.
# Related Work
#### Sentence Representation Learning with NLI
The suitability of leveraging NLI to promote better sentence representation learning is first observed by InferSent, where a siamese BiLSTM network is optimized in a supervised manner with the semantic entailment and contraction classification objective. Universal Sentence Encoder later augments an unsupervised learning objective with the supervised learning on NLI, and shows better transfer performance on various downstream tasks. More recently, SBERT finetunes a siamese BERT model on NLI and sets new state-of-the-art results. However, SBERT as well as the above work adopt the vanilla siamese or triplet loss, which often suffer from slow convergence and bad local optima.
#### Self-Supervised Instance Discrimination
Another relevant line of work is self-supervised contrastive learning, which essentially solves an instance discrimination task that targets at discriminating each positive pair from all negative pairs within each batch of data. Owing to their notable successes, self-supervised instance discrimination has become a prominent pre-training strategy for providing effective representations for a wide range of downstream tasks. While recent successes are primarily driven by the computer vision domain, there is an increasing interest in leveraging variant instance discrimination tasks to support Pretraining Language Models (PLMs). Our proposal can be seen as complementary to this stream of work, considering that the instance discrimination learning based PLMs provide a good foundation for PairSupCon to further enhance the sentence representation quality by further learning on NLI. As demonstrated in Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"}, by training a pre-trained BERT-base model for less than an hour, PairSupCon attains substantial improvement on various downstream tasks that involve sentence semantics understanding at different granularities.
#### Deep Metric Learning
Inspired by the pioneering work of, many recent works have shown significant benefit in learning deep representations using either siamese loss or triplet loss. However, both losses learn from individual pairs or triplets, which often require substantial training data to achieve competitive performance. Two different streams of work have been proposed to tackle this issue, with the shared focus on nontrivial pairs or triplets optimization. propose hard negative or hard positive mining that often requires expensive sampling. extends the vanilla triplet loss by contrasting each positive example against multiple negatives. Our work leverages the strength of both lines, with the key difference being the above work requires categorical level supervision for selecting hard negatives. To be more specific, negative samples that have different categorical labels from the anchor but are currently mapped close to the anchor in the representation space, are likely to be more useful and hence being sampled. However, there are no categorical labels available in NLI. We thereby contrast each positive pair against multiple negatives collected using an unsupervised importance sampling strategy, for which the hypothesis is that hard negatives are more likely to locate close to the anchor. The effectiveness of this assumption is investigated in Section [4.3](#subsec:ablation){reference-type="ref" reference="subsec:ablation"}.
# Model
Following SBERT, we adopt the SNLI and MNLI as our training data, and refer the combined data as NLI for convenience. The NLI data consists of labeled sentence pairs and each can be presented in the form: (*premise, hypothesis, label*). The premise sentences are selected from existing text sources and each premise sentence is paired with variant hypothesis sentences composed by human annotators. Each label indicates the hypothesis type and categorizes semantic relation of the associated premise and hypothesis sentence pair into one of three categories: *entailment, contradiction*, and *neural*, correspondingly. Prior work solely optimizes either siamese loss or triplet loss on NLI. We instead aiming to leverage the implicit grouping effect of instance discrimination learning to better capture the high-level categorical semantic structure of data while, simultaneously promoting better convergence of the low-level semantic textual entailment and contradiction reasoning objective.
## Instance Discrimination
We leverage the *positive (entailment)* pairs of NLI to optimize an instance discrimination objective which tries to separate each positive pair apart from all other sentences. Let \(\mathcal{D} = \left\{(x_j, x_{j'}), y_j\right\}_{j=1}^{M}\) denote a randomly sampled minibatch, with \(y_i= \pm 1\) indicating an entailment or contradiction pair. Then for a premise sentence \(x_i\) within a *positive* pair \((x_i, x_{i'})\), we aim to separate its hypothesis sentence \(x_i'\) from all other \(2M\)-2 sentences within the same batch \(\mathcal{D}\). To be more specific, let \(\mathcal{I} = \left\{i, i'\right\}_{i=1}^{M}\) denote the corresponded indices of the sentence pairs in \(\mathcal{D}\), we then minimize the following for \(x_i\), \[\begin{aligned}
\ell_{\text{ID}}^i =-\log \frac{\exp(s(z_i, z_{i'})/\tau)}{\sum_{j \in \mathcal{I}\setminus i} \exp(s(z_{i}, z_j)/\tau)}\;. \label{eq:pairsupcon_pos} \end{aligned}\] In the above equation, \(z_j = h(\psi(x_j))\) denotes the output of the instance discrimination head in Figure [\[fig:pairsupcon_model\]](#fig:pairsupcon_model){reference-type="ref" reference="fig:pairsupcon_model"}, \(\tau\) denotes the temperature parameter, and \(s(\cdot)\) is chosen as the cosine similarity, *i.e., * \(\text{s}(\cdot) = z_i^Tz_{i'} / \|z_i\| \|z_{i'}\|\). Notice that Equation ([\[eq:pairsupcon_pos\]](#eq:pairsupcon_pos){reference-type="ref" reference="eq:pairsupcon_pos"}) can be interpreted as a (\(2M\)--1)-way softmax based classification loss of classifying \(z_i\) as \(z_i'\). Similarly, for the hypothesis sentence \((x_{i'})\) we also try to discriminate its premise \((x_{i})\) from all the other sentences in \(\mathcal{D}\). We denote the corresponding loss as \(\ell_{\text{ID}}^{i'}\) that is defined by exchanging the roles of instances \(x_{i'}\) and \(x_{i}\) in Equation [\[eq:pairsupcon_pos\]](#eq:pairsupcon_pos){reference-type="eqref" reference="eq:pairsupcon_pos"}, respectively. In summary, the final instance discrimination loss is averaged over all positive pairs in \(\mathcal{D}\), \[\begin{aligned}
\mathcal{L}_{\text{ID}} = \frac{1}{P_M}\sum_{i=1}^{M} \mathbbm{1}_{y_i=1} \cdot \left(\ell_{\text{ID}}^{i} + \ell_{\text{ID}}^{i'}\right)\;. \label{eq:pairsupcon_batchloss_pos} \end{aligned}\] Here, \(\mathbbm{1}_{(\cdot)}\) denotes the indicator function, and \(P_M\) is the number of positive pairs in \(\mathcal{D}\). As demonstrated in Section [4](#sec:experiments){reference-type="ref" reference="sec:experiments"}, optimizing the above loss not only helps implicitly encode categorical semantic structure into representations, but also promotes better pairwise semantic reasoning capability, though no pairwise supervision except the true entailment labels are present to the model.
## Learning from Hard Negatives {#sec:hargneg_sample}
Notice that Eq ([\[eq:pairsupcon_pos\]](#eq:pairsupcon_pos){reference-type="ref" reference="eq:pairsupcon_pos"}) can be rewritten as \[\begin{aligned}
&\ell^i_{\text{ID}} = \nonumber \\ &\log\left(1 + \sum_{j \neq i, i'}\exp\left[\frac{\text{s}(z_{i}, z_j)-\text{s}(z_i, z_{i'})}{\tau}\right]\right)\;. \nonumber \end{aligned}\] It can be interpreted as extending the vanilla triplet loss by treating the other \(2M\)--2 samples within the same minibatch as negatives. However, the negatives are uniformly sampled from the training data, regardless of how informative they are. Ideally, we want to focus on hard negatives that are from different semantic groups but are mapped close to the anchor, *i.e., * \(z_i\), in the representation space. Although the categorical level supervision is not available in NLI, we approximate the importance of the negative samples via the following, \[\begin{aligned}
\label{eq:weighted_instdisc} &\ell^i_{\text{wID}}= \\ & \log\left(1 + \sum_{j \neq i, i'}\exp\left[\frac{\alpha_j\text{s}(z_{i}, z_j)-\text{s}(z_i, z_{i'})}{\tau}\right]\right)\;. \nonumber \end{aligned}\] Here \(\alpha_j = \frac{\exp(\text{s}(z_i, z_{j})/\tau)}{\frac{1}{2M-2}\sum_{k \neq i, i'} \exp(\text{s}(z_{i}, z_k)/\tau)}\), which can be interpreted as the relative importance of \(z_j\) among all \(2M\)-2 negatives of anchor \(z_{i}\). The assumption is that hard negatives are more likely to be those that are located close to the anchor in the representation space. Although there might exists false negatives, *i.e., * those located close to the anchor \(z_i\) but are from the same category, the probability is low as long as the underlying number of categories of the training data is not too small and each minibatch \(\mathcal{D}\) is uniformly sampled.
## Entailment and Contradiction Reasoning
The instance discrimination loss mainly focuses on separating each positive pair apart from the others, whereas there is no explicit force in discriminating contradiction and entailment. To this end, we jointly optimize a pairwise entailment and contradiction reasoning objective. We adopt the softmax-based cross-entropy to form the pairwise classification objective. Let \(u_i=\psi(x_i)\) denote the representation of sentence \(x_i\), then for each labeled sentence pair \((x_i, x_{i'}, y_i)\) we minimize the following, \[\begin{aligned}
\ell_{\text{C}}^i = \text{CE}\left(f(u_i, u_{i'}, |u_i-u_{i'}|), y_i\right)\;. \label{eq:pair_classify} \end{aligned}\] Here \(f\) denotes the linear classification head in Figure [\[fig:motivation_validation\]](#fig:motivation_validation){reference-type="ref" reference="fig:motivation_validation"}, and CE is the cross-entropy loss. Different from, we exclude the neural pairs from the original training set and focus on the binary classification of semantic entailment and contradiction only. Our motivation is that the concept semantic neural can be well captured by the instance discrimination loss. Therefore, we drop the neural pairs from the training data to reduce the functional redundancy of the two losses in PairSupCon and improve the learning efficiency as well.
#### Overall loss
In summary, our overall loss is \[\begin{aligned}
\mathcal{L} = \sum_{i=1}^{M} \ell_{\text{C}}^i + \beta \mathbbm{1}_{y_i=1} \cdot \left(\ell_{\text{wID}}^{i} + \ell_{\text{wID}}^{i'}\right)\;, \end{aligned}\] where \(\ell_{\text{C}}^i\) and \(\ell_{\text{wID}}^{i}, \ell_{\text{wID}}^{i'}\) are defined in Equations ([\[eq:pair_classify\]](#eq:pair_classify){reference-type="ref" reference="eq:pair_classify"}) and ([\[eq:weighted_instdisc\]](#eq:weighted_instdisc){reference-type="ref" reference="eq:weighted_instdisc"}), respectively. In the above equation, \(\beta\) balances between the capability of pairwise semantic entailment and contradiction reasoning, and the capability of the high-level categorical semantic structure encoding. We dive deep into the trade-off between these two aspects by evaluating PairSupCon with different \(\beta\) values in Section [4.3](#subsec:ablation){reference-type="ref" reference="subsec:ablation"}, and show how different \(\beta\) values can benefit different downstream tasks. Unless otherwise specified, we set \(\beta=1\) in this paper for the purpose of providing effective representations to various downstream tasks instead of being tailored to any specific ones.
:::
#### PairSupCon versus. Its Components
In Figure [\[fig:motivation_ablation\]](#fig:motivation_ablation){reference-type="ref" reference="fig:motivation_ablation"}, we compare PairSupCon against its two components, namely Classification and InstanceDisc. As it shows, InstanceDisc itself outperforms Classification on both STS and categorical concept encoding. The result matches our expectation that contrasting each positive pair against multiple negatives, despite obtained through unsupervised sampling, yields better performance than simply learning from each individual pair. By jointly optimizing both objectives, PairSupCon leverages the implicit grouping effect of InstanceDisc to encode the high-level categorical structure into representations while, simultaneously complementing InstanceDisc with more fine-grained semantic concept reasoning capability via Classification. Table [2](#tab:PairSupCon_ablation){reference-type="ref" reference="tab:PairSupCon_ablation"} indicates a trade-off between the high-level semantic structure encoding and the low-level pairwise entailment and contradiction reasoning capability of PairSupCon. Focusing more on the pairwise classification objective, *i.e., * using smaller \(\beta\) values, can hurt the embeddings of the high-level semantic structure. This result is not surprising, especially considering that sentences forming a contradiction pair are not necessarily belong to different semantic groups. On the other hand, InstanceDisc only focuses on separating each positive pair from all other samples within the same minibatch, and an explicit force that discriminates semantic entailment from contradiction is necessary for PairSupCon to achieve competitive performance on the more fine-grained pairwise similarity reasoning on STS. As indicated in Table [2](#tab:PairSupCon_ablation){reference-type="ref" reference="tab:PairSupCon_ablation"}, we can tune the \(\beta\) values to attain effective representations for specific downstream tasks according to their semantic granularity focuses. We set \(\beta=1\) for all our experiments with the goal to provide effective universal sentence representations to different downstream tasks.
#### Hard Negative Sampling Helps
In Table [3](#tab:importance){reference-type="ref" reference="tab:importance"}, we compare both PairSupCon and InstanceDisc against their counterparts where the negatives in the instance discrimination loss are uniformly sampled from data. As it shows, the hard negative sampling approach proposed in Section [3.2](#sec:hargneg_sample){reference-type="ref" reference="sec:hargneg_sample"} leads to improved performance on both STS and clustering tasks. We associate this performance boost with our assumption that hard negatives are likely located close to the anchor. A properly designed distance-based sampling approach can drive the model to better focus on hard negative separation and hence lead to better performance. On the other hand, hard negative sampling without any supervision is a very challenging problem, especially considering that samples within the local region of an anchor are also likely from the same semantic group as the anchor. As a consequence, a solely distance-based sampling approach can induce certain false negatives and hurt the performance. To tackle this issue, leveraging proper structure assumption or domain-specific knowledge could be potential directions, which we leave as future work.
## Transfer Learning {#subsec:transfer_learning}
In order to provide a fair and comprehensive comparison with the existing work, we also evaluate PairSupCon on the following seven transferring tasks: MR, CR, SUBJ, MPQA, SST, TREC, and MRPC. We follow the widely used evaluation protocol, where a logistic regression classifier is trained on top of the frozen representations, and the testing accuracy is used as a measure of the representation quality. We adopt the default configurations of the SentEval toolkit and report the evaluation results in Table [\[tab:senteval_compare_full_supervision\]](#tab:senteval_compare_full_supervision){reference-type="ref" reference="tab:senteval_compare_full_supervision"} in Appendix [\[appendix:transfer_learning\]](#appendix:transfer_learning){reference-type="ref" reference="appendix:transfer_learning"}. As we can see, the performance gap between different methods are relatively small. We suspect the reason is that the transfer learning tasks do not present enough complexities to fully uncover the performance gap between different approaches, especially considering that most tasks are binary classification with a large amount of labeled training examples. To further examine our hypothesis, we extend the evaluation to the setting of few-shot learning, where we uniformly sample 16 labeled instances per class for each task. We report the mean and standard deviation of the evaluation performance over 5 different sample sets in Table [\[tab:senteval_fsl_compare\]](#tab:senteval_fsl_compare){reference-type="ref" reference="tab:senteval_fsl_compare"}. Although we observe more obvious performance gap on each specific task, there is no consistent performance gap between different approaches when evaluated across different tasks. Therefore, to better evaluate the transfer learning performance of sentence representations, more complex and diverse datasets are required.
# Discussion and Conclusion
In this paper, we present a simple framework for universal sentence representation learning. We leverage the implicit grouping effect of instance discrimination learning to better encoding the high-level semantic structure of data into representations while, simultaneously promoting better convergence of the lower-level semantic entailment and contradiction reasoning objective. We substantially advance the previous state-of-the-art results when evaluated on various downstream tasks that involve understanding semantic concepts at different granularities. We carefully study the key components of our model and pinpoint the performance gain contributed by each of them. We observe encouraging performance improvement by using the proposed hard negative sampling strategy. On the other hand, hard negative sampling without any supervision is an crucial, yet significantly challenging problem that should motivate further explorations. Possible directions include making proper structure assumption or leveraging domain-specific knowledge. The substantial performance gain attained by our model also suggests developing explicit grouping objectives could be another direction worth investigation.
[^1]: Please refer to Appendix [\[appendix:cluster_data_stats\]](#appendix:cluster_data_stats){reference-type="ref" reference="appendix:cluster_data_stats"} for more details.
[^2]: We computed the clustering accuracy using the Hungarian algorithm.
[^3]: We randomly select 10 independent seeds and fix them for all evaluations
[^4]: Same as SBERT and SimCSE, we concatenate all the topics and report the overall Spearman's correlation.
[^5]: We also evaluate PairSupCon with STS-Benchmark as the development set in Appendix [\[appendix:sts_dev\]](#appendix:sts_dev){reference-type="ref" reference="appendix:sts_dev"}. | {'timestamp': '2022-02-01T02:18:26', 'yymm': '2109', 'arxiv_id': '2109.05424', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05424'} |
null | null |
null | null |
# Introduction
Recently, synthetic non-Hermitian optical systems based on parity-time (PT) symmetry have attracted considerable attentions due to the unique optical effects they provide, which hold great promise for practical applications not achievable in conventional Hermitian optics. The degeneracies in parameter space of non-Hermitian optical systems, i.e., spectral singularities known as exceptional points, possess unconventional features compared to their Hermitian counterparts, where not only the eigenvalues but also the corresponding eigenvectors of the underlying system coalesce simultaneously. Optical response around exceptional points shows counter-intuitive phenomena, such as, unidirectional invisibility, single-mode lasing, enhanced sensitivity, and enhanced spontaneous emission. For example, dynamically encircling an EP can lead to chiral behaviors, while encircling the EP in different directions results in different output states. Also, a recent study shows that whether or not the dynamics is chiral actually depends on the starting point of the loop. Exceptional points have also been studied in high order or higher dimensions, for example, two-dimensional (2D) exceptional surfaces or 3D exceptional hypersurface. Up to now, EPs have been studied in diverse systems, such as photonic crystals (PCs), plasmonic waveguides, and optomechanical systems. On the other hand, non-Hermiticity can change the band topology in topology systems or even induce a topological phase transition from a trivial system to a topological counterpart. For example, non-Hermitian perturbations can deform a Dirac cone and spawn a ring of exceptional points rather than opening up a gap in Hermitian systems. Bulk Fermi arc and polarization half charge from paired exceptional points have also been observed. A nontrivial Berry phase of the eigenstates can be obtained by encircling exceptional points. Furthermore, edge states around interface between non-Hermitian topological crystals show new features and the effect of gain and loss on quantum-spin-Hall (QSH) edge states has attracted great interest recently . It is interesting to see how gain and loss can affect the trivial and nontrivial edge states in a dramatically different way and most importantly, to observe their difference (trivial and nontrivial) in a single system. In this work, gain and loss are introduced into a dielectric PC with \(C_{6v}\) symmetry that has been proved to mimic the QSH system and show nontrivial topology . We analyze their band structures and topological invariants based on first-principle calculation. Three systems, namely the PT-symmetric, PT-asymmetric, and lossy systems, have been investigated. We provide a comprehensive study on their band structures and topological invariants for both non-Hermitian and Hermitian systems. Moreover, we study the evolution of the trivial and nontrivial edge states under the influence of gain and loss in a single system.
# Bulk Band of non-Hermitian PCs
A 2D PC arranged in a deformed honeycomb lattice with six cylinders in each unit cell is shown in Fig. [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}(a), where the original positions of the cylinders in a honeycomb lattice are indicated by the dashed circles. \(\bm{a_1}\) and \(\bm{a_2}\) are the two translation vectors and \(a_0\) is the lattice constant. It has been shown that expansion of the six cylinders with respect to the center of the hexagonal unit cell results in a nontrivial topology of the PC . In this paper, our aim is to understand how the introduction of gain and loss can lead to nontrivial phase transitions in the topological PC structure. We consider two types of cylinders, A and B, with the same radii but different dielectric constants. Figure [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}(a) shows one type of distributions of A and B. We focus on the transverse-magnetic (TM) modes. The master equation is written as \[\nabla \times \nabla \times \mathbf{E}_z(\mathbf{r})-k_0^2\epsilon(\mathbf{r})\mathbf{E}_z(\mathbf{r})= 0, \label{master}\] where \(k_0\) is the free-space wave number and \(\epsilon(\mathbf{r})\) is the position-dependent dielectric constant. Due to the periodicity of the PC, we can restrict the problem to a single unit cell \[the black rhombus in Fig. [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}(a)\] and as such the field component \(E_z(\mathbf{r})\) could be expressed as Bloch states, \[E_\mathbf{k}(\mathbf{r}) = e^{i \mathbf{k} \cdot \mathbf{r}} u_\mathbf{k}(\mathbf{r}). \label{bloch}\] Here, \(u_\mathbf{k}(\mathbf{r})\) are periodic functions with the same periodicity of the lattice, i.e., \(u_\mathbf{k}(\mathbf{r})=u_\mathbf{k}(\mathbf{r}+\mathbf{R})\) for any lattice vector \(\mathbf{R}\). By applying the generalized finite-difference (FD) method, Eq. ([\[master\]](#master){reference-type="ref" reference="master"}) could be rewritten in a matrix form, \(ME_z =k_0^2E_z\). The eigenequation can then be solved numerically to get the eigenmodes \(E_z\) and eigenvalues \(k_0^2\) . It should be noted that \(M\) is not Hermitian when lossy and gain materials are included. The eigenmodes we defined and calculated are the right eigenmodes, if not mentioned otherwise. When the PC structure is Hermitian (\(\gamma_A=\gamma_B=0\)), the eigenfrequencies of all the bulk bands are real. There is a complete band gap between two doubly degenerate bands at the \(\Gamma\) point \[Fig. [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}(b)\]. When we add gain and loss with the same strength into A and B which are distributed as in Fig. [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}(a), the PC system preserves the PT symmetry. In Fig. [\[band_3case\]](#band_3case){reference-type="ref" reference="band_3case"}, the complex band structures for three typical cases with different \(\gamma\) are plotted. With a small \(\gamma\), as shown in Fig. [\[band_3case\]](#band_3case){reference-type="ref" reference="band_3case"}(a), the band gap remains and the frequencies are still real. The system is in the PT exact phase. At a specific value of \(\gamma\), the bands at the lower and upper edges of the band gap begin to merge. The four modes coalesce into a single mode at the \(\Gamma\) point, forming an exceptional point. As \(\gamma\) increases further, the system enters the PT broken phase: the bulk gap closes as the frequencies of the upper and lower bands become "zipping\" complex conjugate pairs near the \(\Gamma\) point. It should be noticed that from a tight-binding estimation, the critical \(\gamma\) for the PT phase transition is equal to the difference of the inter-and intrasite hoppings, which is roughly \(1.3\) in our case. When the dielectric PC is purely lossy, i.e. \(\gamma_A=\gamma_B>0\), we can see from its band structures in Fig. [\[loss\]](#loss){reference-type="ref" reference="loss"}(a) that the imaginary parts of the frequencies are all positive, indicating that the eigenstates are all lossy. It is known that scaled dielectric function \(\epsilon(\mathbf{r})*s\) results in scaled mode frequencies \(\omega/\sqrt s\). Therefore, the added loss introduces a complex scaling factor so that the imaginary part of the band structure presents the same trend as in the real part. By distributing the cylinders in a different manner, the PT-asymmetric PC is analyzed as shown in Fig. [\[loss\]](#loss){reference-type="ref" reference="loss"}(b). Different from the PT-symmetric PC, the eigenfrequencies become complex as soon as the gain and loss are introduced. Their imaginary parts have positive and negative values but they are not conjugate pairs. In Fig. [\[pt_phase\]](#pt_phase){reference-type="ref" reference="pt_phase"}, we present the evolution of the imaginary parts of the frequencies for the PT-symmetric, PT-asymmetric, and lossy systems as \(\gamma\) increases, where the maximum imaginary part of the four doubly degenerate bands divided by \(\gamma\) is shown. From the results, we can see that the PT-symmetric system exhibits a PT phase transition with a threshold at around \(\gamma=1.3\), while the PT-asymmetric and lossy systems show a thresholdless transition and continuous evolution as \(\gamma\) increases. This phenomenon is consistent with the PT-symmetric quantum spin Hall system .
# Wilson loop characterization
Chern number is used as the topological invariant of non-Hermitian systems with broken time symmetry . For topological systems with spin degree of freedom, the Wilson loop classifies their different topological properties . To calculate the Wilson loop, first, we discretize the first Brillouin zone (BZ) along the reciprocal lattice vectors, \(\bm{b_1}\) and \(\bm{b_2}\), as shown in Fig. [\[WL\]](#WL){reference-type="ref" reference="WL"}(a). The Wilson loop can be calculated at each discretized \(\mathbf{k}\) point, swept from \(\Gamma\) to \(\Gamma'\) or \(\Gamma\) to \(\Gamma''\). For isolated bands, the Wilson loop along \(\bm{b_2}\) is obtained using \[W(\mathbf{k}_{(1,j)})= \operatorname{Im} \left[ \ln \prod_{m=1}^{M-1} U_{\mathbf{k}_{(m,j)} \rightarrow \mathbf{k}_{(m+1,j)}} \right], \\ j=1,2...,N. \label{calWL}\] Here, \(U_{\mathbf{k}_{\alpha} \rightarrow \mathbf{k}_{\beta}} = \frac {\left\langle u_{\mathbf{k}_{\alpha}} | u_{\mathbf{k}_{\beta}}\right\rangle} {\left|\left\langle u_{\mathbf{k}_{\alpha}} | u_{\mathbf{k}_{\beta}} \right\rangle\right|}\), and \(\left\langle u_{\mathbf{k}_{\alpha}} | u_{\mathbf{k}_{\beta}} \right\rangle = \int \epsilon(\mathbf{r}) u_{\mathbf{k}_{\alpha}}^{*} (\mathbf{r}) \cdot u_{\mathbf{k}_{\beta}}(\mathbf{r}) d^2 \mathbf{r}\). \(u_{\mathbf{k}}(\mathbf{r})\) is calculated by using our FD solver. The Wilson loop is gauge invariant. Therefore, in numerical calculation, although an arbitrary phase is added at each calculated eigenstate, it will not alter the results. For example, if we discretize the BZ into four plaquettes, the multiplication in [\[calWL\]](#calWL){reference-type="eqref" reference="calWL"} can be written as \(\left\langle U_1^R|U_2^R\right\rangle \left\langle U_2^R|U_3^R \right\rangle \left\langle U_3^R|U_4^R \right\rangle \left\langle U_4^R|U_5^R \right\rangle\). Here \(R\) means we are calculating the right eigenvectors. Apparently, the states within the BZ show up twice during the multiplication, so their phase factors cancel. Similarly, we can use left eigenvectors, i.e., \(\left\langle U_1^L|U_2^L\right\rangle \left\langle U_2^L|U_3^L \right\rangle \left\langle U_3^L|U_4^L \right\rangle \left\langle U_4^L|U_5^L \right\rangle\). Alternatively, combination of right and left eigenvectors can also be adopted. Since the phases for right and left eigenvectors are different, for the states at the same locations, the same eigenvectors should be chosen. For example, by using \(\left\langle U_1^R|U_2^L\right\rangle \left\langle U_2^L|U_3^R \right\rangle \left\langle U_3^R|U_4^L \right\rangle \left\langle U_4^L|U_5^R \right\rangle\) and \(\left\langle U_1^L|U_2^R\right\rangle \left\langle U_2^R|U_3^L \right\rangle \left\langle U_3^L|U_4^R \right\rangle \left\langle U_4^R|U_5^L \right\rangle\), correct results can also be obtained. These findings are consistent with the claims in . Besides of the phases of the states within the BZ, the phases at the two boundaries should also be taken into consideration. The states at the boundaries are equivalent, i.e. \(E_{\mathbf{k}_{(1,j)}}(\mathbf{r}) = E_{\mathbf{k}_{(M,j)}}(\mathbf{r})\). From [\[bloch\]](#bloch){reference-type="eqref" reference="bloch"}, we have \[\begin{aligned}
\begin{split} E_{\mathbf{k}_{(1,j)}}(\mathbf{r}) &= e^{i \mathbf{k}_{(1,j)} \cdot \mathbf{r}} u_{\mathbf{k}_{(1,j)}}(\mathbf{r}),\\ E_{\mathbf{k}_{(M,j)}}(\mathbf{r}) &= e^{i \mathbf{k}_{(M,j)} \cdot \mathbf{r}} u_{\mathbf{k}_{(M,j)}}(\mathbf{r}). \label{EK} \end{split} \end{aligned}\] Therefore, \[\begin{aligned}
\begin{split} u_{\mathbf{k}_{(M,j)}}(\mathbf{r}) &= E_{\mathbf{k}_{(M,j)}}(\mathbf{r})e^{-i \mathbf{k}_{(M,j)} \cdot \mathbf{r}}\\ &= u_{\mathbf{k}_{(1,j)}}(\mathbf{r})e^{i (\mathbf{k}_{(1,j)}-\mathbf{k}_{(M,j)}) \cdot \mathbf{r}}\\ &= u_{\mathbf{k}_{(1,j)}}(\mathbf{r})e^{-i \bf{b}_2 \cdot \mathbf{r}}, \label{uK} \end{split} \end{aligned}\] which means the states at the boundaries should vary by a fixed phase factor in spite of the arbitrary phases added from the numerical solver. However, when [\[master\]](#master){reference-type="eqref" reference="master"} is numerically solved, the arbitrary phases are different for the boundary states, where we should implement gauge fixing . If we denote the real calculated eigenstate at \(\mathbf{k}_{(M,j)}\) as \(E'_{\mathbf{k}_{(M,j)}}(\mathbf{r})\), then, \[u'_{\mathbf{k}_{(M,j)}}(\mathbf{r}) = E'_{\mathbf{k}_{(M,j)}}(\mathbf{r})e^{-i \mathbf{k}_{(M,j)} \cdot \mathbf{r}}. \label{uK'}\] Combining [\[uK\]](#uK){reference-type="eqref" reference="uK"} with [\[uK\'\]](#uK'){reference-type="eqref" reference="uK'"}, we have \[\begin{aligned}
\begin{split} u_{\mathbf{k}_{(M,j)}}(\mathbf{r}) &= E_{\mathbf{k}_{(M,j)}}(\mathbf{r}) u'_{\mathbf{k}_{(M,j)}}(\mathbf{r}) / E'_{\mathbf{k}_{(M,j)}}(\mathbf{r})\\ &= E_{\mathbf{k}_{(1,j)}}(\mathbf{r}) u'_{\mathbf{k}_{(M,j)}}(\mathbf{r}) / E'_{\mathbf{k}_{(M,j)}}(\mathbf{r}). \end{split} \end{aligned}\] For degenerate bands, the Wilson loop is calculated using \[\begin{aligned}[b] W(\mathbf{k}_{(1,j)}) &= \operatorname{Im} \left[ \ln \prod_{m=1}^{M-1} S_{\mathbf{k}_{(m,j)} \rightarrow \mathbf{k}_{(m+1,j)}} \right], j=1,2...,N;\\ S_{\mathbf{k}_{\alpha} \rightarrow \mathbf{k}_{\beta}} &=\left[ \begin{matrix} {\left\langle u_{\mathbf{k}_{\alpha}}^1 | u_{\mathbf{k}_{\beta}}^1\right\rangle} & {\left\langle u_{\mathbf{k}_{\alpha}}^1 | u_{\mathbf{k}_{\beta}}^2\right\rangle} & \dots & {\left\langle u_{\mathbf{k}_{\alpha}}^1 | u_{\mathbf{k}_{\beta}}^n\right\rangle} \\ {\left\langle u_{\mathbf{k}_{\alpha}}^2 | u_{\mathbf{k}_{\beta}}^1\right\rangle} & {\left\langle u_{\mathbf{k}_{\alpha}}^2 | u_{\mathbf{k}_{\beta}}^2\right\rangle} & \dots & {\left\langle u_{\mathbf{k}_{\alpha}}^2 | u_{\mathbf{k}_{\beta}}^n\right\rangle} \\ \vdots & \vdots & \ddots & \vdots \\ {\left\langle u_{\mathbf{k}_{\alpha}}^n | u_{\mathbf{k}_{\beta}}^1\right\rangle} & {\left\langle u_{\mathbf{k}_{\alpha}}^n | u_{\mathbf{k}_{\beta}}^2\right\rangle} & \dots & {\left\langle u_{\mathbf{k}_{\alpha}}^n | u_{\mathbf{k}_{\beta}}^n\right\rangle} \\ \end{matrix} \right] \label{calWL2} \end{aligned}\] where the superscript \(n\) of \(u^n_{\mathbf{k}}\) indicates the band index. Similarly, a gauge fixing is needed for \(u_{\mathbf{k}_{(M,j)}}\). From Fig. [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}(b), we can see that the first two bands are degenerate at the \(K\) point and the second and third bands are degenerate at the \(\Gamma\) point. Therefore, we need to compute the first three bands together by using [\[calWL2\]](#calWL2){reference-type="eqref" reference="calWL2"}. We find that a discretization of the BZ along \(\bm{b_1}\) into ten plaquettes can guarantee the convergence of the results. The calculated Wilson loops corresponding to the first three bands are shown in the Fig. [\[WL\]](#WL){reference-type="ref" reference="WL"}. For the Hermitian case, the results are symmetric about zero, and the Wannier centers of the Wilson loops are localized at the edge of the unit cell (\(W = \pm \pi\)), which verifies the topological nature. Same feature is observed for the PT-symmetric system with small gain and loss \[middle panel of Fig. [\[WL\]](#WL){reference-type="ref" reference="WL"}(b)\]. When the gain and loss are large enough to close the band gap, the Wilson loop is disturbed from the \(\Gamma\) point, as shown in the right panel in Fig. [\[WL\]](#WL){reference-type="ref" reference="WL"}(b). For the lossy system, as in Fig. [\[WL\]](#WL){reference-type="ref" reference="WL"}(c), the Wilson loops move downward compared with the results in the Hermitian system and the displacement has a positive correlation with the loss. Except for that, the lossy system shows similar properties as the Hermitian system, which implies that the lossy system has almost the same topological feature as the original system. Quite different from the PT-symmetric system, the PT-asymmetric system starts to show different topological feature when small gain and loss are introduced. The Wilson loops are distorted from the \(\Gamma\) point, and are fully out of the original shape when gain and loss are further increased.
# Edge states analysis
The interface between two PCs with different topologies supports topologically protected edge states. In Hermitian systems, the edge states have real eigenfrequencies. When gain and loss are introduced, the edge states are expected to show additional features resulting from the non-Hermiticity of the system . We analyze a topological waveguide that is constructed by introducing an air gap in the topologically nontrivial PC (Fig. [\[cyl\]](#cyl){reference-type="ref" reference="cyl"}) as shown in Fig. [\[edge_band\]](#edge_band){reference-type="ref" reference="edge_band"} . When the system is Hermitian, three bands are identified within the band gap with real eigenfrequencies. The top and bottom bands in Fig. [\[edge_band\]](#edge_band){reference-type="ref" reference="edge_band"}(b), have large group velocities around \(k_x=0\) with states possessing even symmetry along the center of the structure, while the whole middle band has nearly zero group velocities with states possessing odd symmetry. We can see from the distributions of Poynting vectors that the energy flows concentrate at the PC-air interfaces. The net energy flows point to \(-x\) and \(x\) directions for the two states marked by triangles and inverted triangles, in accordance with their slopes at the band structure. Interestingly, the energy flows wind along consecutive half-orbit paths through the dielectric area reversely for the two states. Hence, the two bands represent two pseudospin channels, similar to the QSH effect. In contrast, the odd-symmetric states are just trivial defect states resulting from the symmetry of the structure. When gain and loss increase incrementally, only the eigenfrequencies of top and bottom bands become complex conjugate pairs, starting from \(k_x=0\), while the middle band keeps real \[Fig. [\[edge_band\]](#edge_band){reference-type="ref" reference="edge_band"}(c)\]. The symmetries of the states keep unchanged as in the Hermitian system, but the energy flows change. Nevertheless, the orbital directions of the states of the top and bottom bands along the dielectric area are the same as in Fig. [\[edge_band\]](#edge_band){reference-type="ref" reference="edge_band"}(b). Therefore, in the non-Hermitian system, the two bands of even-symmetric states can be classified based on the pseudospin polarizations and the band of trivial defect states keeps real.
# Conclusion
In conclusion, we have studied the effects of gain and loss on the band structures of a topological PC structure and its interface with air using the generalized FD method . Three cases have been considered, i.e., PT-symmetric, PT-asymmetric and lossy systems. We found that while the PT-asymmetric and lossy systems show no signature of phase transition when increasing the strength \(\gamma\) of gain and loss, the PT-symmetric system shows a phase transition from the PT exact phase to PT broken phase at around \(\gamma=1.3\). Moreover, the method of Wilson loop is adopted to characterize different PC systems. It has been found that the topological feature keeps for PT-symmetric and lossy systems when the band gap is kept. More importantly, the nontrivial edge states in PT-symmetric system show complex conjugate pairs of eigenfrequencies compared to its Hermitian counterparts, while the trivial defect states always show real eigenfrequencies. | {'timestamp': '2022-01-07T02:13:09', 'yymm': '2109', 'arxiv_id': '2109.05498', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05498'} |
null | null |
# Introduction
The Quantum Spectral Curve (QSC) has become an indispensable tool of precision spectroscopy in \(\text{AdS}_5/\text{CFT}_4\) and \(\text{AdS}_4/\text{CFT}_3\) holographic models . For a review on the QSC, see . In this paper, we shall take a step towards extending this powerful method to the spectral problem in another important holographic duality, namely planar \(\text{AdS}_3/\text{CFT}_2\). It is believed that \(\text{AdS}_3/\text{CFT}_2\) dual pairs with 8+8 supersymmetries are integrable .[^1] This is the maximal amount of supersymmetry that is allowed for string theory backgrounds of the form \(\text{AdS}_3\times \mathcal{M}_7\), with \({\cal M}_7=\mathrm{S}^3\times\mathrm{T}^4\) or \(\mathcal{M}_7=\mathrm{S}^3\times\mathrm{S}^3\times\mathrm{S}^1\). The symmetries of these two backgrounds are respectively the small and large \((4,4)\) superconformal symmetries, whose Lie sub-algebras are \(\alg{psu}(1,1|2)^2\) and \(\alg{d}(2,1;\alpha)^2\). The exact S matrices can be found by imposing compatibility with the (centrally extended) vacuum-preserving symmetry algebras of the two theories , much like what can be done in higher-dimensional cases . In this paper, we will focus on string theory on \(\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4\) supported by R-R charge. An important difference between these theories and higher-dimensional integrable string backgrounds is the presence of massless excitations in the worldsheet theory, in addition to the more familiar massive ones. The resulting integrable 2-to-2 S matrix breaks up into independent pieces for the scattering of massless/massless, massive/massive and mixed mass excitations. Expressed in terms of Zhukovsky variables, the S matrices resemble those of higher-dimensional integrable holographic theories, with the mass entering through the shortening conditions. This resemblance is particularly striking in the case of massive excitations , where in the weak-coupling limit the Bethe Equations (BEs) reduce to those of a homogeneous nearest-neighbour \(\alg{psu}(1,1|2)\times \alg{psu}(1,1|2)\) spin-chain, with the two factors only connected by the level-matching condition. Away from the weak-coupling limit, the BEs for each \(\alg{psu}(1,1|2)\) wing bear a striking similarity to the corresponding part of the \(\alg{psu}(2,2|4)\) BEs of \(\text{AdS}_5/\text{CFT}_4\). These observations suggest that (at least a part of) the \(\text{AdS}_3/\text{CFT}_2\) Q-system can be constructed using two sets of \(\alg{psu}(1,1|2)\) Q-functions, one for each wing, and coupling them together in a way that is consistent with the crossing. The Q-system is an important part of any known QSC. In this paper, instead of deriving the QSC following a long route from TBA equations, we use the Q-system as a starting point supplying it with the analyticity properties following closely the previously known cases. However, fairly quickly we realise that one of the analyticity assumptions must be relaxed in our case--namely we no longer assume the square-root type of singularity near the branch points. This new feature is inherently connected with the properties of the *dressing factors* of . Each S matrix block comes with a dressing factor which is not fixed by symmetry requirements. Dressing factors satisfy crossing equations that follow from the Hopf algebra structure of the theory . In the case of string theory on \(\text{AdS}_3\times\mathrm{S}^3\times\mathrm{T}^4\) supported by R-R charge only, dressing factors which solve these crossing equations have been found . There are two independent dressing phases that enter the massive S matrix, corresponding to either scattering excitations in the same \(\alg{psu}(1,1|2)\) wing or in different wings. Their sum is equal to (twice) the Beisert-Eden-Staudacher (BES) phase , while their difference is a new phase, which appears only at the so-called Hernandez-Lopez order . The relative simplicity of this latter factor is related to the fact that boundstates in the theory can only be made from massive constituent excitations from the same \(\alg{psu}(1,1|2)\) wing. As with all solutions of crossing equations, there is potential for CDD ambiguities due to homogeneous solutions of crossing. The absence of such additional factors was demonstrated in , where it was shown that the proposed dressing factors have exactly the required Dorey-Hofman-Maldacena (DHM) double poles and zeros . In the case of massive modes, crossing maps the two \(\alg{psu}(1,1|2)\) wings into one another. This suggest that, as a consequence of crossing, the two copies of the \(\alg{psu}(1,1|2)\) Q-systems should be related by a suitable analytic continuation. Analogous gluing conditions, which can be traced back to crossing, are known to exist in the \(\text{AdS}_5/\text{CFT}_4\) and \(\text{AdS}_4/\text{CFT}_3\) QSC and are needed in addition to the QQ-relations to constrain the system to a closed system of equations, which can be treated analytically in some limits and by means of numerical analysis in general. Furthermore, the simple gluing of the Q-functions can be shown to produce a rather involved expression for the BES dressing phase when considering the large-volume solution. While a number of ingredients for the current construction are borrowed from the known cases, the new type of near-branch point singularity is a crucial novel ingredient. As a test of our proposal we demonstrated how the ABA equations for the massive sector are precisely reproduced in the large-length limit including the dressing phases. In these considerations, we had to make an additional simplifying assumption about the monodromy of \(\mu\)-function in the asymptotic limit, which we have not managed to prove. At the same time, we only reproduced the massive sector equations, which suggests that removing this assumption could revive all the massless degrees of freedom, but we leave this question for future work. Another important direction is to verify the completeness of our system of equations by solving it either numerically as in or in a near BPS limits like in . An intuitive way in which to understand the effect of massless modes is that the massless dispersion relation can be viewed in an approximate sense as the large coupling limit of the massive one, as long as the particle momentum is kept fixed. In the QSC formalism, the coupling usually controls the distance between the cuts in the rescaled spectral parameter \(u/g\). As a result, in the zero mass limit, one might expect this to lead to a number of quadratic cuts collapsing on top of one another. This suggests that, in models with massless modes, the QSC may have a more general singularity structure near the branch points, rather than the conventional square root behaviour seen in higher-dimensional cases. We also expect the analyticity to be simplified in the purely massless sector by employing the pseudo-relativistic variable of . In fact, the assumption of a square-root singularity is over-restrictive in \(\text{AdS}_3\) because it gives rise to a new algebraic constraint on the Q-functions in addition to the QQ-relations. In turn, such a condition collapses the two wings of Q-functions into one, likely leading to drastically simpler analytic properties such as those seen in the Hubbard model , based on a *single* \(su(2|2)\) symmetry. The rest of the paper is organised as follows. In section [2](#sec:data){reference-type="ref" reference="sec:data"}, we collect pre-existing results on integrability for the AdS\(_3\)/CFT\(_2\) duality, which will inspire our conjecture, and describe the algebraic structure of the Q-system for \(psu(1,1|2)\). Section [3](#sec:QSC){reference-type="ref" reference="sec:QSC"} presents our main proposal for the Quantum Spectral Curve, and describes the unique features of these equations as compared to the previous cases. In section [2.1](#sec:ABA){reference-type="ref" reference="sec:ABA"}, we study the large-volume limit of these equations, reproducing precisely the full Asymptotic Bethe Ansatz for massive modes. Finally, we present our conclusions and discuss some future directions. The paper also contains three appendices collecting some notations and technical details. ***Note added:*** The work described here begun before the epidemic. Shortly after the first wave was coming to an end in Europe, we concluded that the large-length limit was incompatible with square-root cuts as described in section [4.3.1](#sec:notquadratic){reference-type="ref" reference="sec:notquadratic"}. During the recent "Integrability in Lower Dimensional AdS/CFT\" online workshop we learnt that Simon Ekhammar and Dima Volin had also independently come to a similar conclusion. We are grateful to Dima and Simon for informing us of their findings and coordinating on the release date of the manuscripts to the arXiv. Motivated by these discussions, we revisited our construction and found that relaxing the branch-cut condition allows for a consistent definition of the QSC together with a large-length limit that reproduces the all-loop massive ABA equations found in the literature. Our proposal for the QSC seems to be fully consistent with the one published simultaneously in. ***Note added in v3:*** In the published version of this article we present a proposal for the QSC, whose large-volume solution involves the Riemann-Hilbert problem ([\[eq:lastRH\]](#eq:lastRH){reference-type="ref" reference="eq:lastRH"}). In section [4.3.2](#sec:dressingrel){reference-type="ref" reference="sec:dressingrel"}, the solution to these equations is written in terms of \(\chi(u,v)\) functions related to the dressing phases proposed in . These functions satisfy the correct discontinuity equations ([\[eq:lastRH\]](#eq:lastRH){reference-type="ref" reference="eq:lastRH"}), but-- upon closer inspection--they have an additional branch point at \(u\sim \infty\). This extra branch point cancels in the full dressing phase of, but its presence in the building block \(\chi^-(u,v)\) is incompatible with the proposed analyticity properties of the QSC. Thus, the claim that our construction reproduces the dressing phases of should be revised.\
After the publication of this paper, a modified proposal for the dressing phases was made in, for which the extra branch point is absent from \(\chi^-(u,v)\). We believe that these modified phases have a very good chance of arising naturally from our QSC in the large volume limit, at least for the massive-massive case,[^2] which would give extra supporting evidence for our construction and at the same time for the conjectured form of the dressing phases. A detailed analysis will be presented elsewhere (for a preliminary discussion see ).\
The remainder of this arXiv version reproduces in full the published article. We emphasize that our proposal for the QSC remains unmodifed, and the only change in the large-volume analysis comes in the explicit form of \(\chi^-\).
# Data on the AdS\({}_3\)/CFT\({}_2\) integrable system {#sec:data}
In this section we assemble together the known facts about the AdS\({}_3\)/CFT\({}_2\) integrable system. This includes the asymptotic Bethe ansatz for massive modes, classical algebraic curve and the \(psu(1,1|2)\) Q-system.
## Asymptotic Bethe Ansatz {#sec:ABA}
The massive Asymptotic Bethe Ansatz (ABA) equations which we will be referring to are those presented in . The symmetry controlling the Bethe equations is \(\alg{psu}(1,1|2)^2\). Each copy of \(\alg{psu}(1,1|2)\) has associated one momentum carrying root and two auxiliary roots. These are called \(x, y_1\) and \(y_3\) for one copy of \(\alg{psu}(1,1|2)\) and \(\bar{x}, y_{\bar{1}}\) and \(y_{\bar{3}}\), respectively, for the other copy. The explicit form of the BEs is: \[\begin{aligned}
1 &=& \prod_{j=1}^{K_2} \frac{y_{1,k}-x_j^+}{y_{1,k}-x_j^-} \prod_{j=1}^{K_{\bar{2}}} \frac{1-\frac{1}{y_{1,k}\bar{x}_j^-}}{1-\frac{1}{y_{1,k}\bar{x}_j^+}},\nonumber\\ \bigg(\frac{x_k^+}{x_k^-}\bigg)^L &=& \prod_{j\neq k}^{K_2} \frac{x_k^+-x_j^-}{x_k^--x_j^+}\frac{1-\frac{1}{x_k^+ x_j^-}}{1-\frac{1}{x_k^-x_j^+}} \sigma^2(x_k,x_j) \prod_{j=1}^{K_1} \frac{x_k^--y_{1,j}}{x_k^+-y_{1,j}} \prod_{j=1}^{K_3} \frac{x_k^--y_{3,j}}{x_k^+-y_{3,j}} \nonumber\\ &&\times \prod_{j=1}^{K_{\bar{2}}} \frac{1-\frac{1}{x_k^+ \bar{x}_j^+}}{1-\frac{1}{x_k^-\bar{x}_j^-}} \frac{1-\frac{1}{x_k^+ \bar{x}_j^-}}{1-\frac{1}{x_k^-\bar{x}_j^+}} \tilde{\sigma}^2(x_k,\bar{x}_j) \prod_{j=1}^{K_{\bar{1}}} \frac{1-\frac{1}{x_k^-y_{\bar{1},j}}}{1-\frac{1}{x_k^+ y_{\bar{1},j}}}\prod_{j=1}^{K_{\bar{3}}} \frac{1-\frac{1}{x_k^-y_{\bar{3},j}}}{1-\frac{1}{x_k^+ y_{\bar{3},j}}},\nonumber\\ \la{ABA1} 1 &=& \prod_{j=1}^{K_2} \frac{y_{3,k}-x_j^+}{y_{3,k}-x_j^-} \prod_{j=1}^{K_{\bar{2}}} \frac{1-\frac{1}{y_{3,k}\bar{x}_j^-}}{1-\frac{1}{y_{3,k}\bar{x}_j^+}}, \end{aligned}\] \[\begin{aligned}
1 &=& \prod_{j=1}^{K_{\bar{2}}} \frac{y_{\bar{1},k}-\bar{x}_j^-}{y_{\bar{1},k}-\bar{x}_j^+} \prod_{j=1}^{K_2} \frac{1-\frac{1}{y_{\bar{1},k}x_j^+}}{1-\frac{1}{y_{\bar{1},k} x_j^-}},\nonumber\\ \bigg(\frac{\bar{x}_k^+}{\bar{x}_k^-}\bigg)^L &=&\prod_{j\neq k}^{K_{\bar{2}}} \frac{\bar{x}_k^--\bar{x}_j^+}{\bar{x}_k^+-\bar{x}_j^-}\frac{1-\frac{1}{\bar{x}_k^+ \bar{x}_j^-}}{1-\frac{1}{\bar{x}_k^-\bar{x}_j^+}} \sigma^2(\bar{x}_k,\bar{x}_j) \prod_{j=1}^{K_{\bar{1}}} \frac{\bar{x}_k^+-y_{\bar{1},j}}{\bar{x}_k^--y_{\bar{1},j}} \prod_{j=1}^{K_{\bar{3}}} \frac{\bar{x}_k^+-y_{\bar{3},j}}{\bar{x}_k^--y_{\bar{3},j}}\nonumber\\ &&\times \prod_{j=1}^{K_2} \frac{1-\frac{1}{\bar{x}_k^-x_j^-}}{1-\frac{1}{\bar{x}_k^+ x_j^+}} \frac{1-\frac{1}{\bar{x}_k^+ x_j^-}}{1-\frac{1}{\bar{x}_k^-x_j^+}} \tilde{\sigma}^2(\bar{x}_k,x_j) \prod_{j=1}^{K_1} \frac{1-\frac{1}{\bar{x}_k^+ y_{1,j}}}{1-\frac{1}{\bar{x}_k^-y_{1,j}}}\prod_{j=1}^{K_3} \frac{1-\frac{1}{\bar{x}_k^+ y_{3,j}}}{1-\frac{1}{\bar{x}_k^-y_{3,j}}},\nonumber\\ 1 &=& \prod_{j=1}^{K_{\bar{2}}} \frac{y_{\bar{3},k}-\bar{x}_j^-}{y_{\bar{3},k}-\bar{x}_j^+} \prod_{j=1}^{K_2} \frac{1-\frac{1}{y_{\bar{3},k}x_j^+}}{1-\frac{1}{y_{\bar{3},k} x_j^-}}. \end{aligned}\] The Bethe equations are written in the grading illustrated in Fig. [\[fig:dyn\]](#fig:dyn){reference-type="ref" reference="fig:dyn"}. The massless modes will not be included in our analysis, and they do not feature anywhere in the Bethe equations we write. There is a further level-matching constraint on the solutions to the Bethe equations, in the form of \[\begin{aligned}
1 = \prod_{j=1}^{K_2} \frac{x_j^+}{x_j^-} \prod_{j=1}^{K_{\bar{2}}} \frac{\bar{x}_j^+}{\bar{x}_j^-} \end{aligned}\] (once more disregarding massless modes). The Zhukovsky variables satisfy the familiar constraint given by (suppressing the particle index) \[x^+ + \frac{1}{x^+}-x^--\frac{1}{x^-} = \frac{i}{h}, \qquad \frac{x^+}{x^-} = e^{ip},\] where \(h\) is the coupling constant of the theory and \(p\) is the particle momentum. The same holds for the barred variables. The dispersion relation that gives the energy of a particle of momentum \(p\) reads \[E(p) = \sqrt{1 +16 \, h^2 \sin^2 \frac{p}{2}},\] and the anomalous dimension of the state associated to a solution of the ABA is given by \[\delta\Delta \equiv \gamma = 2 h \sum_{k=1}^{K_2}\left(\frac{i}{x_{k}^+}-\frac{i}{x_{k}^-}\right) + 2 h \sum_{j=1}^{K_{\bar{2}}}\left(\frac{i}{\bar{x}_j^+}-\frac{i}{\bar{x}_{j}^-}\right).\] The explicit form of the dressing phases from is given by \[\sigma(p_1,p_2) = \mathrm{e}^{ i \theta(p_1,p_2)}, \qquad \tilde{\sigma}(p_1,p_2) = \mathrm{e}^{ i \tilde{\theta}(p_1,p_2)},\] with the familiar splitting \[\theta(p_1,p_2) = \chi(x_1^+,x_2^+)+\chi(x_1^-,x_2^-)-\chi(x_1^+,x_2^-)-\chi(x_1^-,x_2^+),\] with similar expressions for \(\tilde{\sigma}\). The individual blocks read \[\begin{aligned}
\la{chidef} && \chi(x,y) = \chi^{\mbox{\scriptsize BES}}(x,y)-\frac{1}{2}\Big[\chi^{\mbox{\scriptsize HL}}(x,y)-\chi^-(x,y)\Big], \\ \nonumber&& \tilde{\chi}(x,y) = \chi^{\mbox{\scriptsize BES}}(x,y)-\frac{1}{2}\Big[\chi^{\mbox{\scriptsize HL}}(x,y)+\chi^-(x,y)\Big]. \end{aligned}\] The part denoted by BES is the Beisert-Eden-Staudacher dressing phase-its expression can be found reproduced in the review . The same holds for the HL part, referring to the Hernandez-Lopez phase \[\begin{aligned}
\chi^{{\rm HL}}(x,y) = \Bigg(\int_{C^+}-\int_{C^-}\Bigg) \frac{dw}{4\pi} \frac{1}{x-w} \Big[ \log (y-w)-\log(y-\frac{1}{w})\Big]. \end{aligned}\] The new ingredient which was constructed in is given by \[\begin{aligned}
\chi^-(x,y) = \Bigg(\int_{C^+}-\int_{C^-}\Bigg) \frac{dw}{8\pi} \frac{1}{x-w} \log \Big[ (y-w)\Big(1-\frac{1}{y w}\Big)\Big]-x \leftrightarrow y, \end{aligned}\] where the contours \(C^\pm\) denote the upper (resp., lower) half semicircle in the complex \(w\)-plane, both running anti-clockwise. These expressions are valid in the physical region \(|x|>1, |y|>1\). The notation \(\chi^-\) is commonly used in the \(\text{AdS}_3\) literature for this portion of the phase. The minus sign should not be confused with a shift in the spectral parameter-as will otherwise always be meant in this paper. Since we will be merely concerned with the massive modes, it is expected that the Asymptotic Bethe equations which we have written above should be valid exactly in the coupling \(h\) but only asymptotically in the length \(L\). In other words, wrapping corrections are expected to be exponentially suppressed . This situation would be rather different were we to include massless modes, whose impact on the TBA is not exponentially suppressed-they are expected to be polynomially suppressed in the presence of mixed massive-massless interactions , or require exact solutions as in the case of the conformal TBA of (see also ). Notice also that \(4h\) gives the size of the branch cut which goes to zero at weak coupling. Since all interaction between the two \(\alg{psu}(1,1|2)\) wings go through the branch-cut, the two wings become completely decoupled in the limit of small coupling constant \(h\rightarrow 0\), except for the level-matching condition.
## Main features of the classical curve {#sec:classical}
The Quantum Spectral Curve is a quantum version of the classical curve, which thus contains crucial structural hints. We shall from now on denote with un-dotted/dotted indices the variables pertaining to the first/second wing, respectively, of the Dynkin diagram--corresponding to the first/second copy of \(\alg{psu}(1,1|2)\). Here we present a short description of some aspects of the algebraic curve describing the integrability of classical solutions of string theory on \(AdS_3\times S^3 \times T^4\), following the discussion in . This description is based on 4+4 quasimomenta, associated to the fundamental representations of the two \(\alg{psu}(1,1|2)\)'s. They will be denoted by \((p_1^A, p_2^A, p_1^{S},p_2^{S})\) and \((p_{\dot 1}^A, p_{\dot 2}^A, p_{\dot 1}^{S},p_{\dot 2}^{S})\). Each quasimomentum naturally parametrises motion in one of the factors of the target space, which is marked by the superscripts \(A\), \(S\) for \(AdS_3\) and \(S^3\), respectively. They are very important quantities which are expected to arise in a WKB-type approximation of the Q-functions in the classical limit of the quantum spectral curve. The \(p\)'s are naturally seen as functions of the Zhukovsky variables, and contain the symmetry charges of the solution in their asymptotics: (
)&&(
)=(
),\
(
)&&(
)=(
), where on the rhs we used the explicit expression of the charges in terms of Bethe roots numbers. Finally, the classical curve tells us how the quasimomenta in the two wings are related. In particular, for the quasimomenta describing motion in \(AdS_3\), the relation is extremely simple and consists in analytic continuation \[\la{clasglue} p_a^A\left(\frac{1}{x}\right) = p_{\dot a}^A(x), \;\; a=1,2\,,\] as described in equations (7.13) and (7.38) of . We will lift this property to the quantum case.
## Algebra of the \(\alg{psu}(1,1|2)\) Q-system {#sec:Qsystem}
The sets of functional relations between the Q-functions (known as Q-systems) take a universal form depending only on the symmetry algebra of the integrable system. Since our model contains two copies of \(\alg{psu}(1,1|2)\), important input for our construction comes from the structure of QQ relations for this algebra. The \(\alg{psu}(1,1|2)\) Q-system contains \(16\) independent Q-functions depending on the spectral parameter \(u\). They can be labelled as \(Q_{A|I}\), where \(A\), \(I\) are completely anti-symmetric strings of indices made from \(\left\{1,2\right\}\) \[A\,, I \in \left\{ \emptyset, 1, 2, (12) \right\}\,,\] interrelated by the QQ relations [\[eq:QQrel\]]{#eq:QQrel label="eq:QQrel"} Q\_a A\|I Q\_A\|I i&=& Q\_aA\| I i\^+ Q\_A\|I \^--Q\_aA\|Ii\^-Q\_A\|I\^+ , [\[eq:fermionic\]]{#eq:fermionic label="eq:fermionic"}\
Q\_12\|I Q\_ \|I &=& Q\_1\|I\^+ Q\_2\|I\^--Q\_1\|I\^-Q\_2\|I\^+ ,\
Q\_A\|12 Q\_ A\| &=& Q\_A\|1\^+ Q\_A\|2\^--Q\_A\|1\^-Q\_A\|2\^+ ,[\[eq:bosonic2\]]{#eq:bosonic2 label="eq:bosonic2"} where \(a,i \in\left\{1,2\right\}\) are single indices, and \(A\), \(I\) are anti-symmetric multi-indices defined above. The first type of relation ([\[eq:fermionic\]](#eq:fermionic){reference-type="ref" reference="eq:fermionic"}) is usually called fermionic, and the remaining two bosonic. In these equations, we are using the notation adopted in the whole paper for shifts in the spectral parameter \(u\): for any function \(g\), \[g^{[\pm n]}(u)\equiv g(u\pm \tfrac{i n}{2})\,, \qquad\qquad g^{\pm}(u)\equiv g^{[\pm 1]}(u).\] In our proposal, the QSC will contain two copies of these relations, which we will denote by distinguishing between dotted and undotted indices (giving us 16+16 Q-functions). In this section, we focus on one wing, and elaborate on some consequences of ([\[eq:fermionic\]](#eq:fermionic){reference-type="ref" reference="eq:fermionic"})-([\[eq:bosonic2\]](#eq:bosonic2){reference-type="ref" reference="eq:bosonic2"}). We will make a simple special choice for the Q-functions with the extremal combinations of indices: \[\label{eq:unimodular} Q_{\emptyset|\emptyset} = Q_{12|12} = 1,\] which is analogous to the choice made in the other known QSC cases. Notice that the Q-system has several symmetries, and in particular we are free to set \(Q_{\emptyset|\emptyset} = 1\) through an overall normalisation. The further, nontrivial algebraic assumption underlying ([\[eq:unimodular\]](#eq:unimodular){reference-type="ref" reference="eq:unimodular"}) is that \(Q_{12|12}(u)/Q_{\emptyset|\emptyset}(u)\) is an \(i\)-periodic function of \(u\). Once we have this periodicity property, the analytic properties of Q-functions we will discuss in the next sections imply that \(Q_{12|12}/Q_{\emptyset|\emptyset}\) should be a constant, which we are free to normalise to one using the symmetries of the Q-system. In quantum spin chains, the periodicity \(1=\frac{Q_{12|12}^+ Q_{\emptyset|\emptyset}^-}{Q_{12|12}^-Q_{\emptyset|\emptyset}^+ }\) can be traced to the quantum transfer matrix having a unit determinant. It is also expected that such condition reflects the projectivity of the algebra \(\alg{psu}(1,1|2)\). In particular, as discussed in, it implements a zero-charge constraint for the quantum numbers, which enter the asymptotics of Q-functions in the way described in the next section. For these reasons, from now on we assume the validity of ([\[eq:unimodular\]](#eq:unimodular){reference-type="ref" reference="eq:unimodular"}), which so far seems fully consistent with the description of the \(AdS_3\) integrable system. We will adopt a special notation for some of the Q-functions, \[{\bf Q}_k \equiv Q_{\emptyset|k}, \;\;\; {\bf P}_a\equiv Q_{a|\emptyset}, \;\;\; {\bf Q}^k \equiv \epsilon^{kl} Q_{12|l}, \;\;\;{\bf P}^a \equiv \epsilon^{ab} Q_{b|12},\] as well as \(Q^{a|i}\equiv \epsilon^{ab} \epsilon^{ij} Q_{b|j}\). Explicitly, \[\label{eq:Qaiupexpl} Q^{a|i} = \left( \begin{array}{cc} Q_{2|2} &-Q_{2|1}\\-Q_{1|2} & Q_{1|1} \end{array}\right),\] such that \[Q_{a|i} Q^{b|i} = \delta_a^b, \;\; \;\;Q_{a|i} Q^{a|j} = \delta_i^j,\] due to the unimodularity property \[\text{det}\left( Q_{a|i} \right) = 1,\] which is a consequence of the Q-system with the boundary conditions ([\[eq:unimodular\]](#eq:unimodular){reference-type="ref" reference="eq:unimodular"}). Let us write explicitly some of the fermionic equations, which will be used extensively, \[\label{eq:fermionic1} Q_{a|i}^+-Q_{a|i}^-= {\bf P}_a {\bf Q}_i,\] together with \(Q_{a|i}^--Q_{a|i}^+ = Q_{12|i} Q_{a|12}\), which can be rewritten in Hodge-dual notation as \[\label{eq:fermionic1dual} Q^{a|i \;+}-Q^{a|i\;-} =-{\bf P}^a {\bf Q}^i.\] Further useful consequences of the QQ relations are:[^3] \[\label{eq:vanishPP} {\bf P}_a {\bf P}^a = {\bf Q}_i {\bf Q}^i = 0,\] and the following relations \[\label{eq:raisewithQai} Q_{a|i}^{\pm} {\bf Q}^i = {\bf P}_a \;\;, \;\; Q_{a|i}^{\pm} {\bf P}^a = {\bf Q}_i \;\;,\;\; Q^{a|i\;\pm}{\bf Q}_i={\bf P}^a\;\;,\;\; Q^{a|i\;\pm}{\bf P}_a={\bf Q}^i\;,\] where the equations with \(\pm\) signs are compatible due to ([\[eq:fermionic1\]](#eq:fermionic1){reference-type="ref" reference="eq:fermionic1"})--([\[eq:vanishPP\]](#eq:vanishPP){reference-type="ref" reference="eq:vanishPP"}). A useful rewriting of ([\[eq:fermionic1\]](#eq:fermionic1){reference-type="ref" reference="eq:fermionic1"}), ([\[eq:fermionic1dual\]](#eq:fermionic1dual){reference-type="ref" reference="eq:fermionic1dual"}) incorporating \(Q_{a|i}\) is [\[eq:keytranslation\]]{#eq:keytranslation label="eq:keytranslation"} Q\_a\|i\^-= Q\_a\|j\^+ ( \^j_i-**Q**\^j **Q**\_i ), Q\^a\|i- = Q\^a\|j+ ( \_j\^i +**Q**\^i **Q**\_j ), or alternatively, [\[eq:keytranslation2\]]{#eq:keytranslation2 label="eq:keytranslation2"} Q\_a\|i\^-= Q\_b\|i\^+ ( \^b_a-**P**\^b **P**\_a ), Q\^a\|i- = Q\^b\|i+ ( \_a\^b +**P**\^b **P**\_a ). So far, most of these relations are structurally similar to the ones found for \(\alg{psu}(2,2|4)\)-the \(AdS_5\) case. In this case of lower rank, however, there is an interesting new feature, which follows from the fact that \(Q_{a|i}\) and \(Q^{a|i}\) are related in a simple manner by ([\[eq:Qaiupexpl\]](#eq:Qaiupexpl){reference-type="ref" reference="eq:Qaiupexpl"}). The compatibility of ([\[eq:fermionic1\]](#eq:fermionic1){reference-type="ref" reference="eq:fermionic1"}) and ([\[eq:fermionic1dual\]](#eq:fermionic1dual){reference-type="ref" reference="eq:fermionic1dual"}) then gives \[{\bf Q}^k{\bf P}^a=-\epsilon^{kl}\epsilon^{ab}{\bf Q}_l{\bf P}_b,\] or explicitly, \[{\bf Q}^1{\bf P}^1=-{\bf Q}_2{\bf P}_2\;\;,\;\; {\bf Q}^1{\bf P}^2=+{\bf Q}_2{\bf P}_1\;\;,\;\; {\bf Q}^2{\bf P}^1=+{\bf Q}_1{\bf P}_2\;\;,\;\; {\bf Q}^2{\bf P}^2=+{\bf Q}_1{\bf P}_1,\] which imply the equalities of certain ratios of \({\bf P}\) or \({\bf Q}\) functions: \[\la{defchi} \frac{{\bf Q}^1}{{\bf Q}_2}=-\frac{{\bf Q}^2}{{\bf Q}_1} =-\frac{{\bf P}_2}{{\bf P}^1}=+\frac{{\bf P}_1}{{\bf P}^2} \equiv r.\] The quantity \(r(u)\) defined above will have an interesting role in our system. Notice that it allows to raise or lower the indices \[\la{chi} {\bf Q}^k=+r \epsilon^{kl}{\bf Q}_l\;\;,\;\; {\bf Q}_k=-\frac{1}{r} \epsilon_{kl}{\bf Q}^l\;\;,\;\; {\bf P}^k=-\frac{1}{r} \epsilon^{kl}{\bf P}_l\;\;,\;\; {\bf P}_k=+r \epsilon_{kl}{\bf P}^l\;.\] Finally, a useful consequence of the Q-system is the existence of a 2nd order finite difference equation, describing the \({\bf Q}\) functions in terms of the \({\bf P}\) functions (and vice versa). These Baxter-type equations are described in appendix [\[app:Baxter\]](#app:Baxter){reference-type="ref" reference="app:Baxter"}.
#### Q-system and Bethe ansatz.
An important consequence of a Q-system is that it immediately implies the existence of Bethe-like equations restricting the positions of the zeros of the Q-functions, which play the role of Bethe roots. In this argument, we anticipate a crucial assumption on the Q-functions, namely that they do not have any poles. One such system of Bethe equations constrains the zeros of the Q-functions \[Q_{\emptyset|1} = {\bf Q}_1, \;\; Q_{1|1}, \;\; Q_{12|1} =-{\bf Q}^2.\] For instance, from ([\[eq:fermionic1dual\]](#eq:fermionic1dual){reference-type="ref" reference="eq:fermionic1dual"}) we learn that \[\left. Q_{1|1}^+-Q_{1|1}^-\right|_{u \in \left\{ \text{zeros of } {\bf Q}_1 \right\} } = \left. {\bf P}_1 {\bf Q}_1 \right|_{u \in \left\{ \text{zeros of } {\bf Q}_1 \right\} } = 0,\] while, since \({\bf Q}_1 {\bf P}_1 =-{\bf P}^2 {\bf Q}^2\), it is also true that \[\left. Q_{1|1}^+-Q_{1|1}^-\right|_{u \in \left\{ \text{zeros of } {\bf P}_1 \right\} } = \left. Q_{1|1}^+-Q_{1|1}^-\right|_{u \in \left\{ \text{zeros of } {\bf P}^2 \right\} }= \left. Q_{1|1}^+-Q_{1|1}^-\right|_{u \in \left\{ \text{zeros of } {\bf Q}^2 \right\} } = 0.\] Shifting the bosonic equation \(Q_{1|1}^+ Q_{2|1}^--Q_{1|1}^-Q_{2|1}^+ =-{\bf Q}^2 {\bf Q}_1\) by \(\pm i/2\), we also obtain . Q\_1\|1\^++ Q\_2\|1 \|\_u { Q\_1\|1 } &=& .-**Q**\^2 +**Q**\_1\^+ \|\_u { Q\_1\|1 } ,[\[eq:toratio1\]]{#eq:toratio1 label="eq:toratio1"}\
. Q\_1\|1\^-- Q\_2\|1 \|\_u { Q\_1\|1 } &=& . + **Q**\^2-**Q**\_1\^- \|\_u { Q\_1\|1 } .[\[eq:toratio2\]]{#eq:toratio2 label="eq:toratio2"} The above constraints can be recast as the exact Bethe equations[^4] [\[eq:BAasyL\]]{#eq:BAasyL label="eq:BAasyL"} &&. \|\_u { **Q**\_1 } = 1 [\[eq:ABA1L\]]{#eq:ABA1L label="eq:ABA1L"}\
&&. \|\_u { Q\_1\|1 } =-1, [\[eq:ABA2L\]]{#eq:ABA2L label="eq:ABA2L"}\
&&. \|\_u { **Q**\^2 } = 1, [\[eq:ABA3L\]]{#eq:ABA3L label="eq:ABA3L"} where the middle relation comes from the ratio of ([\[eq:toratio1\]](#eq:toratio1){reference-type="ref" reference="eq:toratio1"}),([\[eq:toratio2\]](#eq:toratio2){reference-type="ref" reference="eq:toratio2"}). In a similar way one can deduce several other systems of Bethe equations. For instance, relations of the same form are valid for the zeros of the functions \({\bf P}_1\), \(Q_{1|1}\), \({\bf P}^2\). We write these relations with a dot, anticipating that they will be relevant for the second wing: [\[eq:BAasyR\]]{#eq:BAasyR label="eq:BAasyR"} &&. \|\_u { **P**\_ } = 1 [\[eq:ABA1Lb\]]{#eq:ABA1Lb label="eq:ABA1Lb"}\
&& . \|\_u { Q\_\| } =-1, [\[eq:ABA2Lb\]]{#eq:ABA2Lb label="eq:ABA2Lb"}\
&&. \|\_u { **P**\^ } = 1, [\[eq:ABA3Lb\]]{#eq:ABA3Lb label="eq:ABA3Lb"} In a system like the ones arising in AdS/CFT, the Q-functions are in general complicated functions not known explicitly, therefore such exact Bethe equations have limited practical usefulness when analysing generic solutions of the QSC. However, for certain classes of solutions, such as those with large charges or near special points in the moduli space of the holographic theory the Q-functions do simplify. In the last section of the paper, we find the explicit large-volume limit of some Q-functions, arising from our QSC equations. Exact Bethe equations such as the ones given above will then reduce to the ABA equations. Additionally, AdS\({}_3\)/CFT\({}_2\) dual pairs have multiple moduli, which preserve integrability and at special points in the moduli space of each holographic pair additional simplifications to the exact Bethe equations may occur. For example, the weakly coupled RR-charged theory is expected to describe a nearest-neighbour integrable spin chain .
# Proposal for the QSC {#sec:QSC}
In this section we describe the structure of the proposed Quantum Spectral Curve for \(\text{AdS}_3\). In the absence of the general TBA equations we cannot follow the usual route of to derive the QSC from TBA. Instead we will be guided by the common properties of the known QSCs for \(\text{AdS}_5\) and \(\text{AdS}_4\). If we summarise the known QSCs there are two main ingredients: QQ-relations, and analytical properties of Q-functions. We consider these components in turn in the following.
## Introducing the Q-functions
#### QQ-relations.
In the known case, the QQ-relations follow from the structure of the symmetry of the system. In \(\text{AdS}_3\) we have two copies of \(\alg{psu}(1,1|2)\) and a natural assumption would be to have two copies of QQ-relations for \(\alg{psu}(1,1|2)\), described in the previous section. To distinguish the two copies we will use dotted indices for one of them, so we will use the following sets of indices (\(a=1,2,\;k=1,2\) and same for dotted indices) \_k,\_a,Q\_a\|k && Q\^a\|k,\^k,\^a,\
**Q**\_k,\_a,Q\_a\|k && Q\^a\|k,\^k,\^a. The above Q-functions are related by the QQ-relations. A distinguished subset of them, from which one can recover the remaining Q-functions are \[{\bf P}_a\;,\;{\bf P}^a\;\;{\rm and}\;\; {\bf P}_{\dot a}\;,\;{\bf P}^{\dot a}\;\;{\rm constrained\; by}\;\;{\bf P}_a {\bf P}^a={\bf P}_{\dot a}{\bf P}^{\dot a}=0\;.\] For example, \({\bf Q}_k\) can be reconstructed from \({\bf P}_a\) and \({\bf P}^a\) by solving the second order finite-difference equation \[\la{Bax} {\bf Q}_k^{++} D_1^--{\bf Q}_k D_2 + {\bf Q}^{--}_k D_1^+=0\;,\] with the coefficients depending solely on \({\bf P}\)'s: \[\la{Ds} D_1= \epsilon_{ab} {\bf P}^{a-}{\bf P}^{b+}\;\;,\;\; D_2= \epsilon_{ab} {\bf P}^{a--}{\bf P}^{b++}-{\bf P}_c {\bf P}^{c--}\; \epsilon_{ab} {\bf P}^{a}{\bf P}^{b++}\;.\] The above relation, derived in appendix [\[app:Baxter\]](#app:Baxter){reference-type="ref" reference="app:Baxter"}, is a consequence of the QQ-relations, so an identical equation holds for the dotted Q-functions. Equally one can interchange \({\bf Q}\leftrightarrow{\bf P}\) in and .
#### Classical correspondence.
In the classical limit, described by strong coupling \(h\rightarrow \infty\) and large quantum numbers scaling as \(\sim h\), we expect that the quasimomenta appear in a WKB approximation of some of the Q-functions. In particular, they should be directly related to the Q-functions living in the fundamental representation of each \(\alg{psu}(1,1|2)\) algebra. With the notation borrowed from the other cases, we link \({\bf P}\)'s with the quasi-momenta associated with \(\mathrm{S}^3\) and \({\bf Q}\)'s with the ones for \(\text{AdS}_3\). For the first wing, we will take this correspondence to be the following: [\[eq:classical1\]]{#eq:classical1 label="eq:classical1"} &&( **Q**\_1, **Q**\_2 \| **P**\_1 ,**P**\_2 ) \~( e\^-\^u p\^A_1 , e\^-\^u p\^A_2 \| e\^ + \^u p\^S_1 , e\^+ \^u p\^S_2 ) ,\
&& ( **Q**\^1, **Q**\^2 \| **P**\^1, **P**\^2 ) \~( e\^ \^u p\^A_1 , e\^ \^u p\^A_2 \| e\^-\^u p\^S\_1 , e\^-\^u p\^S_2 ), which is structurally the same as in \(\text{AdS}_5\). For the second wing, we take[^5] [\[eq:classical2\]]{#eq:classical2 label="eq:classical2"} &&( **Q**\_, **Q**\_ \| **P**\_, **P**\_ ) \~( e\^-\^u p\^A_2 , e\^-\^u p\^A_1 \| e\^ \^u p\^S_2 , e\^\^u p\^S_1 ),\
&&( **Q**\^, **Q**\^ \| **P**\^, **P**\^ ) \~( e\^ \^u p\^A\_ , e\^ \^u p\^A\_ \| e\^-\^u p\^S\_ , e\^-\^u p\^S\_ ). [\[eq:classical4\]]{#eq:classical4 label="eq:classical4"}
#### Large-\(u\) asymptotics.
Consistently with the quasi-classical identifications and the asymptotics of the quasimomenta described in section [2.2](#sec:classical){reference-type="ref" reference="sec:classical"}, the Q-functions should exhibit power-law asymptotics at large \(u\), with behaviour characterised by the charges. In particular, we assume \[\label{eq:largeu} {\bf P}_a\simeq A_a u^{M_a}, \;\;\; {\bf P}^a \simeq A^a \, u^{-M_a-1}, \;\;\;{\bf Q}_i\simeq B_i u^{\hat{M}_i}, \;\;\; {\bf Q}^i\simeq B^i u^{-\hat{M}_i-1},\] for large \(u\), where \[\begin{aligned}
M_a&\equiv \left(-\frac{L}{2} + K_2-K_1-1, \;\frac{L}{2}-K_2 + K_3 \right), & \hat{M}_k &\equiv \left(\frac{\gamma}{2} + \frac{L}{2} + K_1 ,\;-\frac{\gamma}{2} -\frac{L}{2}-K_3-1 \right), \end{aligned}\] \[\begin{aligned}
M_{\dot a}&\equiv \left(-\frac{L}{2} + K_{\dot 3}, \;\frac{L}{2}-K_{\dot 1}-1 \right), & \hat{M}_{\dot k} &\equiv \left(\frac{\gamma}{2} + \frac{L}{2} + K_{\dot 2}-K_{\dot 3}-1,\;-\frac{\gamma}{2}-\frac{L}{2}-K_{\dot 2} + K_{\dot 1} \right). \end{aligned}\] In the following sections, we will see that some of the Q-functions have horizontal cuts connecting to infinity. In this case, the asymptotic behaviour above will be assumed to be valid for \(\text{Im}(u)>0\). Notice that the classical identification is valid in a regime of large quantum numbers, so that it only fixes the structure of ([\[eq:largeu\]](#eq:largeu){reference-type="ref" reference="eq:largeu"}) up to finite shifts. However, those can be fine-tuned by the match with the ABA which will be described in the last section of the paper. We will take the exact asymptotics of the Q-functions to be as above.
#### Constraints on the constant prefactors and shortening conditions.
The pre-factors \(A\) and \(B\) in \({\bf P}\) and \({\bf Q}\) functions usually play an important role. They can be determined by plugging the large \(u\) expansion into the QQ-relations or Baxter equation. This leads to the following identities \[r_0\;h^{\hat B}\frac{\prod(-y_{\dot 3,i})}{\prod(-y_{\dot 1,i})}=\frac{B^1}{B_2}=\frac{A_1}{A^2}\,\qquad \quad r_0\;h^{\check B}\frac{\prod(-y_{ 3,i})}{\prod(-y_{ 1,i})}=\frac{B^{\dot 1}}{B_{\dot2}}=\frac{A_{\dot1}}{A^{\dot2}}\;.\] The Baxter equation then implies B_1 B\^1=-B_2 B\^2=,\
A_1 A\^1=-A_2 A\^2=, and with dots B\_ B\^=-B\_ B\^=,\
A\_ A\^=-A\_ A\^=. Above we used the following relation between the charges and the Bethe root numbers: \[\begin{aligned} \Delta &=\gamma+L+K_{\dot{2}} +\frac{K_1}{2}+\frac{K_3}{2}-\frac{K_{\dot{1}}}{2}-\frac{K_{\dot{3}}}{2 }\;, \\ S&=\frac{K_1}{2}+\frac{K_3}{2}+\frac{K_{\dot{1}}}{2}+\frac{K_{\dot{3}}}{2}-K_{\dot{2}}\;, \\ K&=\frac{K_1}{2}+\frac{K_3}{2}+\frac{K_{\dot{1}}}{2}+\frac{K_{\dot{3}}}{2}-K_2\;, \\ J&=L-K_2+\frac{K_1}{2}+\frac{K_3}{2}-\frac{K_{\dot{1}}}{2}-\frac{K_{\dot{3}}}{2}\;, \\ \hat{B}&=K_1-K_3\;, \\ \check{B}&=K_{\dot{1}}-K_{\dot{3}}\;. \end{aligned}\] The half-BPS shortening condition \(\Delta=J\) and \(S=K\) follows from requiring for \(A\) and \(B\) to vanish. This is an integrability-based derivation of a non-renormalization result for theories with small \((4,4)\) super-conformal symmetry. In such theories, there are left or right sub-algebra shortening conditions: \(\Delta_L=J_L\) or \(\Delta_R=J_R\). It is well-known that at generic points in the moduli space states which are short with respect to only one such sub-algebra (*i.e.* quarter-BPS states) are not protected, while states which satisfy both shortening conditions (half-BPS states) do not receive quantum corrections . An independent derivation of these results was found using ABA methods which are valid in the large \(L\) limit. The QSC derivation presented here, showing that only half-BPS states are protected, is valid for all lengths \(L\).
## Analytic properties
As in all other studied cases, we assume that all \(4\) types of \({\bf P}\)'s have only one branch cut \((-2h,2h)\) on the real axis and no other singularities on either sheet of their Riemann surface, as shown on Figure [\[fig:pq\]](#fig:pq){reference-type="ref" reference="fig:pq"}. Since the \({\bf Q}\)-functions are determined in terms of \({\bf P}\)'s by means of equation , the analytic properties of \({\bf Q}\) can be deduced from those of \({\bf P}\). Before describing them let us introduce two different bases of solutions of \(\eq{Bax}\): \_k\^&-&\
**Q**\_k\^&-&. As the coefficients of only have a few cuts near the real axis, and are analytic otherwise, we can always find two solutions of \({\bf Q}_k^\downarrow\) which do not have cuts in the UHP, and another pair of solutions \({\bf Q}_k^\uparrow\) which are analytic in the LHP. Rewriting as \[\la{Bax2} {\bf Q}_k^\downarrow =\frac{{\bf Q}^{\downarrow++}_k {\color{orange}D^{++}_2}-{\bf Q}_k^{\downarrow[+4]} {\color{orange} D_1^+} }{D_1^{[+3]}} \;,\] and assuming that \({\bf Q}_k^\downarrow\) is analytic for \({\rm Im}\;u>0\) we see that the highlighted terms in the r.h.s. will produce a branch cut on the real axis. Iterating further with shifts \(u\to u-2 i n\) in general we generate a ladder of cuts going down the complex plane like on Figure [\[fig:pq\]](#fig:pq){reference-type="ref" reference="fig:pq"}. At the same time, since there are only two linearly independent (with periodic coefficients) solutions of a second order equation there must exist an \(i\)-periodic function (with short cuts) \(\Omega_{k}^{\;\;l}\) which relates the two sets of solutions \[\label{eq:defOmega} {{\bf Q}}^\uparrow_k = \Omega_k^{\;\;m} {\bf Q}^\downarrow_{m}\;\;,\;\; \Omega_k^{\;\;m}(u+i) = \Omega_k^{\;\;m}(u)\;.\] In fact one can write \(\Omega_k^{\; m}\) explicitly in terms of \({\bf Q}\)'s \[\Omega_k^{\;\; m}=\epsilon^{ml}\frac{ {\bf Q}^{\uparrow}_k {\bf Q}^{\downarrow--}_l-{\bf Q}^{\uparrow--}_k {\bf Q}^{\downarrow}_l }{ {\bf Q}^{\downarrow}_1 {\bf Q}^{\downarrow--}_2-{\bf Q}^{\downarrow--}_1 {\bf Q}^{\downarrow}_2 }\;\] and the periodicity can be verified using . There are identical equations for the dotted indices. Furthermore, in Appendix [\[app:Baxter\]](#app:Baxter){reference-type="ref" reference="app:Baxter"} we show that the Hodge-dual Q-functions also satisfy \[\label{eq:upOmega} {\bf Q}^{\uparrow k} = \Omega^k_{\;\;m} {\bf Q}^{\downarrow m}\,\qquad \Omega^k_{\;\;m}\Omega_k^{\;\;l}=\delta^l_m\;.\]
#### Gluing conditions.
So far, the two Q-systems were existing independently. Here we propose a particular way of joining them together. The underlying idea is to fix the apparent asymmetry between the analytic properties of \({\bf Q}\) and \({\bf P}\) (see Figure [\[fig:pq\]](#fig:pq){reference-type="ref" reference="fig:pq"}). Whereas \({\bf P}\) has only one branch-cut, as we argued above, \({\bf Q}\) should have a ladder of cuts going either up or down from the real axis. Following the observation in other QSCs, we notice that a section of the Riemann surface of \({\bf Q}\)'s with long cut i.e. \((-\infty,-2h)\cup(2h,\infty)\) on the real axis should not have any other cuts. More specifically we require that (see Fig. [\[fig:my_label\]](#fig:my_label){reference-type="ref" reference="fig:my_label"}) \[\la{gluing1} {\bf Q}_k^\downarrow(u+i0)=G_k^{\;\;\dot n}\;{\bf Q}_{\dot n}^\uparrow(u-i0)\;\;,\;\; {\bf Q}_{\dot k}^\downarrow(u+i0)=G_{\dot k}^{\;\; n}\;{\bf Q}_{n}^\uparrow(u-i0)\;\;,\;\;u\in (-2h,2h)\] where \(G_k^{\;\;\dot n}\) and \(G_{\dot k}^{\;\;n}\) are two different independent constant matrices. In the studied cases of QSC they have several zero components, but in our case their exact form is still to be deduced. However, one can make a first guess by looking at the classical counterpart of the gluing relations . Using the identification we see that it suggests \(G_{1}^{\;\dot 2}\) and \(G_{2}^{\;\dot 1}\) to be the only non-zero elements of \(G_{k}^{\;\;\dot n}\). For the Hodge-dual Q-functions, the gluing conditions take a similar form \[\la{gluing2} {\bf Q}^{k\downarrow}(u+i0)=G^k_{\;\;\dot n}\;{\bf Q}^{\dot n\uparrow}(u-i0)\;\;,\;\; {\bf Q}^{\dot k\downarrow}(u+i0)=G^{\dot k}_{\;\; n}\;{\bf Q}^{n\uparrow}(u-i0)\;\;,\;\;u\in (-2h,2h).\] Like in the known cases, we assert that gluing is a symmetry of the Q-system \[G^k_{\;\;\dot n} = \epsilon^{kl}\epsilon_{\dot n\dot m} G^{ \;\;\dot m}_{l}\;\;,\;\; G^{\dot k}_{\;\;n} = \epsilon^{\dot k\dot l}\epsilon_{n m} G^{\;\; m}_{\dot l}\;\;,\;\;\det G=1\;.\] In the following, we will choose a basis of Q-functions with specified large-\(u\) asymptotics on the first sheet, described in . after this choice is made, we are not free to diagonalise the gluing matrix with a linear transformation. For this reason, we will keep track of it explicitly throughout. We leave for future work the discussion of the matrix structure of \(G\) in this special basis, but as we argued above the classical limit suggests an off-diagonal structure for this matrix.
#### Properties of the \(r\)-function.
The \(r\)-function, which was defined in section [2.3](#sec:Qsystem){reference-type="ref" reference="sec:Qsystem"} and allows to lower and raise indices, has interesting analyticity properties. From we note that \(r={\bf P}^1/{\bf P}_2\), meaning that \(r\) (and \(\dot r\)) has at most one cut on the main sheet \((-2h,2h)\). At the same time \(r={\bf Q}^\downarrow_1/{\bf Q}^{2\downarrow}\), meaning that it has only one long cut at the same time. In other words the analytic continuation from above \(r^\gamma\) is analytic in the LHP. \[\la{chi} ({\bf Q}^{k\downarrow})(u+i0)=r(u+i0) \epsilon^{kl}{\bf Q}_l^{\downarrow}(u+i0) =r(u+i0) \epsilon^{kl}G_{l}^{\;\;\dot n}{\bf Q}_{\dot n}^\uparrow(u-i0),\] at the same time the l.h.s. can be expressed as \[G^{k}_{\;\; \dot n}\;{\bf Q}^{\dot n\uparrow}(u-i0) = G^{k}_{\;\; \dot n}\;\epsilon^{\dot n\dot m}\dot r(u-i0){\bf Q}_{\dot m}^{\uparrow}(u-i0) = \dot r(u-i0) \epsilon^{kl} G^{\;\; \dot n}_{l} {\bf Q}_{\dot n}^{\uparrow}(u-i0),\] from where we deduce that \(r(u+i0)=\dot r(u-i0)\). Similarly, we can start from the dotted version of the derivation above to get \(r(u-i0)=\dot r(u+i0)\). From this consideration we see that \(r(u)\) has a single quadratic cut, which connects it to \(\dot r(u)\). This branch cut can be rationalised with the help of the Zhukovsky variable \(x(u)=\frac{u+\sqrt{u-2h}\sqrt{u+2h}}{2h}\) so we can write \(r\) explicitly in terms of its zeros/poles[^6] r(u)=r_0, and \(\dot r(u)\) is \(r(u)\) with \(x(u)\) replaced by \(1/x(u)\). \(r_0\) is a constant. In the above expression we assume \(|y_{\dots }|\ge 1\). Finding such a simple expression for a combination of \({\bf P}\)'s is an interesting novel feature of the \(\text{AdS}_3\) QSC.
## On analytic continuation
We now deduce several consequences of the discussion in the previous section. We will see that the simple set of constraints given above implies the existence of a rich mathematical structure. The Q-functions live on a Riemann surface with infinitely many sheets, but the equations we will now deduce allow us to map any one of these sheets to the first one, as is the case also for the other examples of QSCs. As anticipated in the introduction, it will turn out that the branch cuts in this system of QSC equations *cannot be* quadratic. This means that, for any branch point on the Riemann surface, we can go around it in two ways, and in principle this yields two different results. We will introduce the analytic continuation paths \(\gamma\) and its inverse \(\gamma^{-1}\), which we will also denote by \(\bar{\gamma}\). The path \(\gamma\) goes around a branch point at \(2 h\) in anticlockwise sense, or alternatively, it goes around the branch point at \(-2 h\) in clockwise sense. Since in this section we think in terms of short cuts for all the Q-functions, we can say that \(\gamma\) goes through the short cut \((-2h, +2h)\) from above, while \(\bar{\gamma}\) crosses it from below. The two paths are represented in figure [\[fig:gamma\]](#fig:gamma){reference-type="ref" reference="fig:gamma"}. We denote the analytic continuation of any function of \(u\) along these paths as \(f^{\gamma}\) or \(f^{\gamma^{-1}}\equiv f^{\bar{\gamma}}\). In this notation, ([\[gluing1\]](#gluing1){reference-type="ref" reference="gluing1"}) and ([\[gluing2\]](#gluing2){reference-type="ref" reference="gluing2"}) become \[\label{eq:Qtildebar2} ({{\bf Q}}_k^\downarrow)^{\gamma} = G_k^{\;\; \dot m} {\bf Q}^\uparrow_{\dot n} \;\;, \;\; ({{\bf Q}}_{\dot k}^\downarrow)^{\gamma} = G_{\dot k}^{\;\; n} {\bf Q}_{n}^\uparrow, \;\;\;\; ({{\bf Q}}^{k\downarrow} )^{\gamma} = G^k_{\;\; \dot n} \bar{{\bf Q}}^{\dot n\uparrow}, \;\;\;\; ({{\bf Q}}^{\dot n \downarrow})^{\gamma} = G^{\dot a}_{\;\; b} {\bf Q}^{ b\uparrow}.\]
### The \({\bf Q}\omega\)-system
By defining \(i\)-periodic functions \(\omega\): \[\label{eq:om-G-Om} \omega_k^{\;\;\dot n} = G_k^{\;\;\dot m} \Omega_{\dot m}^{\;\; \dot n \,,\qquad\qquad \omega_{\dot k}^{\;\; n} = G_{\dot k}^{\;\; m} \Omega_{ m}^{\;\; n} \,,\] where \(\Omega\)'s are the matrices relating LHPA and UHPA bases in ([\[eq:upOmega\]](#eq:upOmega){reference-type="ref" reference="eq:upOmega"}), ([\[eq:defOmega\]](#eq:defOmega){reference-type="ref" reference="eq:defOmega"}), the system of equations ([\[eq:Qtildebar2\]](#eq:Qtildebar2){reference-type="ref" reference="eq:Qtildebar2"}) can be conveniently rewritten in the form \[\label{eq:Qomega} {\bf Q}^{\downarrow\gamma}_k = \omega_k^{\;\;\dot m} {\bf Q}^\downarrow_{\dot m \,,\qquad\qquad{\bf Q}_{\dot k}^{\downarrow\gamma} = \omega_{\dot k}^{\;\; m} {\bf Q}^\downarrow_{ m}.\] Notice also that by construction, it follows from the properties of the gluing matrix and \(\Omega\) function that \[\omega_k^{\;\;\dot m} \omega^l_{\;\;\dot m} = \delta_k^l \,,\qquad\qquad\text{det}(\omega) = 1.\] Similarly, one can introduce \[\bar\omega_{\dot n}^{\;\; m} = G^k_{\;\;\dot n}\Omega_{\;\; k}^{m},\] such that \[\label{eq:Qomega2} {\bf Q}^{\uparrow\gamma^{-1}}_{\dot k}= \bar\omega_{\dot k}^{\;\; m} {\bf Q}^\uparrow_{ m}.\] In what follows, we adopt a simplified notation[^7], where \({\bf Q}^\downarrow\) is denoted by \({\bf Q}\) and \({\bf Q}^\uparrow\) is denoted by \(\bar {\bf Q}\). So and become \[\label{eq:Qtil} {\bf Q}_k^\gamma = \omega_k^{\;\;\dot m}{\bf Q}_{\dot m}\,,\qquad\qquad \bar{\bf Q}_k^{\bar\gamma} = \bar\omega_k^{\;\;\dot m}\bar{\bf Q}_{\dot m}\;.\] Now let us understand the analytic continuation under the cuts of \(\omega\), focusing on \(\omega_{\dot k}^{\;\;l}\) first. Notice that the matrix \(\Omega_k^{\;\;l}\) can be expressed as \(\Omega_k^{\;\;l} = \bar{Q}_{a|k}^+ Q^{a|l\;+}\) (see ) and since \(Q^{a|i\; +}\) has no cut on the real axis, we only need to understand the analytic continuation of \(\bar{Q}_{a|k}^+\). The defining relation of this function is \[\bar Q_{a|k}^+-\bar Q_{a|k}^-= {\bf P}_a \bar {\bf Q}_k,\] where \(\bar Q_{a|k}^-\) is now analytic and invariant under the analytic continuation along \(\bar\gamma\). Computing the discontinuity we obtain \[\bar{Q}_{a|k}^{+\bar\gamma}-\bar{Q}_{a|k}^+ = {\bf P}_a^{\bar\gamma} \bar{{\bf Q}}_k^{\bar\gamma}-{\bf P}_a \bar{{\bf Q}}_k \;,\] which, multiplied by \(Q^{a|l\;+}\) on the left, leads to [\[eq:Omegatilde\]]{#eq:Omegatilde label="eq:Omegatilde"} (\_k\^l)\^\|-\_k\^l = \|**Q**\_k \^\| **Q**\^l\|-\|**Q**\_k **Q**\^l. Next, multiplying these equations by \(G_{\dot k}^{\;\;k}\), and using ([\[eq:Qtildebar2\]](#eq:Qtildebar2){reference-type="ref" reference="eq:Qtildebar2"}), we find: [\[eq:omegatil\]]{#eq:omegatil label="eq:omegatil"} (\_k\^l)\^\|-\_k\^l = **Q**\_k **Q**\^l\|-**Q**\^\_k **Q**\^l. This expression generalises a similar relation found in \(AdS_5\) and \(AdS_4\) cases, but now we distinguish two different directions for the analytic continuation on the r.h.s.. As usual one can replace dotted to undotted indices to get a similar identity for \(\omega_{ k}^{\;\;\dot l}\). We can use to determine the double continuation of \({\bf Q}_{\dot k}\) along the contour \(\gamma\)--we will then see explicitly that there may be an obstruction to the cuts being quadratic. We start by continuing ([\[eq:Qtil\]](#eq:Qtil){reference-type="ref" reference="eq:Qtil"}) along the inverse path \(\bar{\gamma}\), which gives \[\la{derog} {\bf Q}_{\dot k}=(\omega_{\dot k}^{\;\;l})^{\bar\gamma}{\bf Q}^{\bar\gamma}_l = \omega_{\dot k}^{\;\;l}{\bf Q}^{\bar\gamma}_l-{{\bf Q}}^\gamma_{\dot k} {\bf Q}^l{\bf Q}^{\bar\gamma}_l = \omega_{\dot k}^{\;\;l}\(\delta_l^p-{{\bf Q}}_{l} {\bf Q}^p\){\bf Q}^{\bar\gamma}_p,\] where the second equality is obtained by using ([\[eq:omegatil\]](#eq:omegatil){reference-type="ref" reference="eq:omegatil"}), and recalling that \({\bf Q}_i{\bf Q}^i = 0\). Inverting the factor on the r.h.s., we get \[\label{eq:Qbargamma} {\bf Q}^{\bar\gamma}_p= \(\delta_p^l + {{\bf Q}}_{p} {\bf Q}^l\) \omega^{\dot k}_{\;\;l}{\bf Q}_{\dot k}\;.\] From this we can compute directly the difference of the analytic continuation of the Q-function along \(\gamma\) and \(\bar{\gamma}\): \[{\bf Q}_k^{\gamma}-{\bf Q}_k^{\bar{\gamma}} = {\bf Q}_{\dot m} \left(\omega_k^{\;\;\dot m}-\omega^{\dot m}_{\;\;\; k} \right)+ {\bf Q}_k {\bf Q}^l \omega_l^{\;\;\dot m} {\bf Q}_{\dot m}.\] In the case of \(AdS_5\), the two terms on the r.h.s. would vanish separately, due to the symmetry properties of the analogue of \(\omega\), ensuring that the branch cuts are quadratic. In our case, that does not need to be the case, since \(\omega\) connects different kinds of indices and there is no reason *a priori* to expect any symmetry between them. We make a further interesting observation by rewriting ([\[eq:omegatil\]](#eq:omegatil){reference-type="ref" reference="eq:omegatil"}) in the form [\[eq:omegatil2\]]{#eq:omegatil2 label="eq:omegatil2"} (\_k\^l)\^\|-**Q**\_k **Q**\^l\| = \_k\^l-**Q**\^\_k **Q**\^l ̇This shows immediately that the combination \(\omega_{\dot k}^{\;\;l}-{{\bf Q}}^\gamma_{\dot k} {\bf Q}^l\;\) is equal to its analytic continuation, and therefore the cut on the real axis disappears in this combination. We can also write it as \(\omega_{\dot k}^{\;\;m} (\delta_m^l-{{\bf Q}}_{m} {\bf Q}^l)\). Then taking along \(\gamma\), we get \[\label{eq:usefulgamma} {\bf Q}_{\dot k}^\gamma = \omega_{\dot k}^{\;\;l}\(\delta_l^p-{{\bf Q}}_{l} {\bf Q}^p\){\bf Q}_p = \omega_{\dot k}^{\;\;l}{\bf Q}_l\] with the final equality being in agreement with [\[eq:om-G-Om\]](#eq:om-G-Om){reference-type="eqref" reference="eq:om-G-Om"}. The first equality allows us to find the expression for \({\bf Q}_{\dot k}\) continued a second time along \(\gamma\): \[\la{doubletrouble} {\bf Q}_{\dot k}^{\gamma^2} = \omega_{\dot k}^{\;\;l}\(\delta_l^p-{{\bf Q}}_{l} {\bf Q}^p\){\bf Q}^\gamma_p = \omega_{\dot k}^{\;\;l}\(\delta_l^p-{{\bf Q}}_{l} {\bf Q}^p\)\omega_p^{\;\;\dot h}{\bf Q}_{\dot h}\.\] This expression confirms the potential obstruction to the cuts being quadratic. In particular, we can repeatedly iterate this continuation and obtain in general &&(**Q**\_k )\^\^n =U\_k\^ p (**Q**\_p )\^\^n-1 ,(**Q**\_ k )\^\^n = U\_k\^ p (**Q**\_p )\^\^n-1,\
&&(**Q**\_k )\^\|\^n =\|U\_k\^ p (**Q**\_p )\^\|\^n-1 ,(**Q**\_k )\^\|\^n = \_k\^ p (**Q**\_p )\^\|\^n-1, where && U\_k\^p\_k\^l\(\delta_l^p-{{\bf Q}}_{l} {\bf Q}^p\) ,\|U\_k\^p( U\^-1 )\_k\^p = \(\delta_{\dot k}^{\dot m} + {{\bf Q}}_{\dot k} {\bf Q}^{\dot m}\) \^p\_m ,\
&& \_ k\^p\_ k\^m\(\delta_{\dot m}^{\dot p}-{{\bf Q}}_{\dot m} {\bf Q}^{\dot p}\), \_ k\^p( U\^-1 )\_ k\^p = \(\delta_{ k}^{m } + {{\bf Q}}_{ k} {\bf Q}^{ m}\) \^p\_ m. In general, following the path \(\gamma^n\) produces a concatenation of monodromies \(U \cdot \dot U \cdot U \cdot \dot U \dots\), but since there is no reason to expect \((U \cdot \dot U )\) to be the identity matrix (or a root of the latter), this is nontrivial, meaning that each branch point has infinite order and connects to infinitely many sheets. Notice that, while in general we expect the branch points to be non-quadratic, there are some special combinations of Q-functions that do exhibit this property. We already showed that this is the case for the ratio \(r\) defined in ([\[defchi\]](#defchi){reference-type="ref" reference="defchi"}). We now consider \[{\bf Q}^l{\bf Q}^{\bar\gamma}_l= {\bf Q}^l \omega^{\dot k}_{\;\;l}{\bf Q}_{\dot k}= {\bf Q}^{\dot k\;\gamma}{\bf Q}_{\dot k}\;,\] where we used ([\[eq:Qbargamma\]](#eq:Qbargamma){reference-type="ref" reference="eq:Qbargamma"}) and the analogous equation to ([\[eq:Qtil\]](#eq:Qtil){reference-type="ref" reference="eq:Qtil"}) with (raised, dotted) indices. Lowering the indices with ([\[defchi\]](#defchi){reference-type="ref" reference="defchi"}), and remembering that, as deduced above, \(r^{\gamma} = r^{\bar{\gamma}} =\dot r\), the same relations (and their dotted version) can be written as \[\label{eq:aboveeq} \epsilon^{kl}{\bf Q}_k{\bf Q}^{\bar\gamma}_l=-\epsilon^{\dot k\dot l}{\bf Q}_{\dot k}{\bf Q}^{\gamma}_{\dot l}\,,\qquad\qquad \epsilon^{kl}{\bf Q}_k{\bf Q}^{\gamma}_l=-\epsilon^{\dot k\dot l}{\bf Q}_{\dot k}{\bf Q}^{\bar\gamma}_{\dot l}\;.\] Continuing the first equation above along \(\gamma\), we get \[\label{eq:beloweq} \epsilon^{kl}{\bf Q}_k^\gamma{\bf Q}_l=-\epsilon^{\dot k\dot l}{\bf Q}_{\dot k}^\gamma{\bf Q}^{\gamma^2}_{\dot l} \;,\] but due to the second equation in ([\[eq:aboveeq\]](#eq:aboveeq){reference-type="ref" reference="eq:aboveeq"}), the l.h.s. is also equal to \(-\epsilon^{\dot k\dot l}{\bf Q}_{\dot k}^{\bar\gamma}{\bf Q}_{\dot l}\), meaning that the combination \(\epsilon^{\dot k\dot l}{\bf Q}_{\dot k}^{\bar\gamma}{\bf Q}_{\dot l}= \epsilon^{kl}{\bf Q}_k{\bf Q}^{\gamma}_l\) comes back after \(\gamma^2\)! As a final observation, we notice that, continuing the two sides of [\[eq:omegatil2\]](#eq:omegatil2){reference-type="eqref" reference="eq:omegatil2"} along \(\gamma\), one can also obtain an explicit equation for \(\omega^{\gamma}\) in terms of quantities on the first sheet. The main results of this section can be summarised in the following equations:[^8] \[({\bf Q})^{\gamma}_{\dot k} = \omega_{\dot k}^{\;\; l} {\bf Q}_l ,\;\; \;\;\;\; ({\bf Q}^{\dot k} )^{\gamma} = \omega^{\dot k}_{\;\; l} {\bf Q}^l ,\label{eq:Qomegafirst}\] ,and \[\left((\omega)^{\bar{\gamma}}-\omega\right)_{\dot k}^{\;\; l}={\bf Q}_{\dot k} ({\bf Q}^{ l})^{\bar{\gamma}}-({\bf Q}_{\dot k})^{\gamma} {\bf Q}^{ l}, \;\;\;\; \left((\omega)^{\bar{\gamma}}-\omega\right)^{\dot k}_{\;\; l}=-{\bf Q}^{\dot k} ({\bf Q}_{ l})^{\bar{\gamma}} + ({\bf Q}^{\dot k})^{\gamma} {\bf Q}_{l}. \label{eq:Qomegalast}\] Here, as usual, we understand that for every equation there is its double obtained by interchanging dotted and undotted indices. Together with \({\bf Q}_i{\bf Q}^i = 0\), \(\omega_{\dot k}^{\; l} \omega^{\dot k}_{\; m }= \delta_m^l\), and the periodicity of \(\omega\), the relations ([\[eq:Qomegafirst\]](#eq:Qomegafirst){reference-type="ref" reference="eq:Qomegafirst"}),([\[eq:Qomegalast\]](#eq:Qomegalast){reference-type="ref" reference="eq:Qomegalast"}) may be taken as a self-consistent description of the QSC, which is usually dubbed \({\bf Q}\omega\)-system.[^9] Bouncing back and forth between these equations, and using the fact that \(\omega\) is \(i\)-periodic, one can obtain the result of any analytic continuation of the \({\bf Q}\)-functions and \(\omega\) functions, inside any cut, and express it in terms of their values on the first sheet. This is the same feature that was observed in the other examples of QSC, see the discussion in. It is encouraging that this property is still valid here, even though the analytic structure is more complicated due to the branch points having infinite order.
### The \({\bf P}\mu\)-system
We now describe the constraints on the analytic continuation of \({\bf P}\) functions. Analogously to , the main object in this case is the matrix \(\mu\) defined as \[\label{eq:defmu} \mu_a^{\; \; \dot b} \equiv Q_{a|c}^-\, \omega^c_{\;\;\dot d} \,Q^{\dot b|\dot d \;\;-} \,,\qquad \mu^a_{\;\; \dot b} \equiv Q^{a|c\;-} \, \omega_c^{\;\;\dot d}\, Q_{\dot b|\dot d }^{-},\] which will play a role similar to \(\omega\). Notice that just like in the case of \(\omega\), \(\mu\) has unit determinant and \(\mu_a^{\;\;\dot b} \mu^c_{\;\;\dot b} = \delta_a^c\). We also notice the alternative expression \[\label{eq:barmu} \mu_{a}^{\;\;\dot b} = \bar{Q}_{a|i}^-( G_{\dot k}^{\;\; i} \omega^{\dot k}_{\;\; l} G^l_{\;\;\dot m} ) \bar{Q}^{\dot b| \dot m \;-},\] which is obtained through the relations ([\[eq:QupQdown\]](#eq:QupQdown){reference-type="ref" reference="eq:QupQdown"}), and will become useful in the discussion of the next section. While \(\omega\) is an \(i\)-periodic function on the Riemann section with short cuts, \(\mu\) has a periodicity on the section with long cuts, as depicted in figure [\[fig:mu\]](#fig:mu){reference-type="ref" reference="fig:mu"}. Expressed in terms of a section with short cuts, this "mirror periodicity" becomes \[\label{eq:mirrorperiodic} \mu^{++} = (\mu)^{\gamma}.\] To prove this relation (and thus also long-cut periodicity of \(\mu\)), we continue it along \(\bar{\gamma}\) and show that the combination \((\mu^{++} )^{\bar{\gamma}}-\mu\) vanishes. We can rewrite such a difference as ((\^++ )\^\|-)\_a\^b = Q\_a\|i\^+(\^i \_k )\^\| Q\^b\|k +-Q\_a\|i\^-\^i\_k Q\^b\|k -. We can now plug in \((\omega)^{\gamma^{-1}}\) from the (undotted version of) ([\[eq:Qomegalast\]](#eq:Qomegalast){reference-type="ref" reference="eq:Qomegalast"}), and, in the second term, use the identities ([\[eq:keytranslation\]](#eq:keytranslation){reference-type="ref" reference="eq:keytranslation"}) to relate \(Q_{a|i}^-\) and \(Q_{a|i}^+\). We get &&Q\_a\|k\^+( \^k \_m + (**Q**\^k)\^ **Q**\_m-**Q**\^k ( **Q**\_m )\^\| ) Q\^b \|m +\
&& - Q\_a\|k\^+ ( \^k_m-**Q**\^k **Q**\_m ) \^m\_m (\^m\_k + **Q**\^m **Q**\_k ) Q\^b\|k + = 0, where a perfect cancellation occurs due to ([\[eq:Qomegafirst\]](#eq:Qomegafirst){reference-type="ref" reference="eq:Qomegafirst"}),([\[eq:Qomegalast\]](#eq:Qomegalast){reference-type="ref" reference="eq:Qomegalast"}), establishing ([\[eq:mirrorperiodic\]](#eq:mirrorperiodic){reference-type="ref" reference="eq:mirrorperiodic"}). We can use \(\mu\) to compute the values of \({\bf P}\) on the second sheet. In particular, the definition ([\[eq:defmu\]](#eq:defmu){reference-type="ref" reference="eq:defmu"}), together with \({\bf P}_a = Q_{a|i}^+ {\bf Q}^i\), immediately implies \[\label{eq:Ptil0} ({\bf P}_a)^{\gamma} =\mu_a^{\;\dot b \;++} {\bf P}_{\dot b} =(\mu_a^{\;\dot b })^{\gamma} {\bf P}_{\dot b},\] which is conveniently rewritten as \[\label{eq:Ptil} ({\bf P}_a )^{\bar{\gamma}} = {\bf P}_{\dot b} \mu^{\dot b}_{\;\; a} \;.\] This equation, compared to ([\[eq:Qtil\]](#eq:Qtil){reference-type="ref" reference="eq:Qtil"}), highlights the symmetry of the construction between \({\bf P}\) and \({\bf Q}\) functions. From ([\[eq:omegatil\]](#eq:omegatil){reference-type="ref" reference="eq:omegatil"}), it is also immediate to derive \[(\mu_a^{\;\;\dot b})^{\gamma}-\mu_a^{\;\;\dot b} = Q_{a|k}^+\left( \omega^k_{\;\;\dot m}-(\omega^k_{\;\;\dot m})^{\bar{\gamma}} \right) Q^{\dot b|\dot m+} = {\bf P}_a ({\bf P}^{\dot b} )^{\bar{\gamma}}-({\bf P}_a)^\gamma {\bf P}^{\dot b},\] which shows that the combination \(\mu_a^{\;\;\dot b} + {\bf P}_a ({\bf P}^{\dot b} )^{\bar{\gamma}} = \left(\delta_a^b + {\bf P}_a {\bf P}^b\right)\mu_b^{\;\;\dot b}\) has no cut on the real axis. From this observation and ([\[eq:Ptil\]](#eq:Ptil){reference-type="ref" reference="eq:Ptil"}) we also deduce \[({\bf P}_a)^{\gamma} = \left(\delta_a^b + {\bf P}_a {\bf P}^b\right)\mu_b^{\;\;\dot b}{\bf P}_{\dot b}\,,\qquad\qquad ({\bf P}_a)^{\gamma^2} = \left(\delta_a^b + {\bf P}_a {\bf P}^b\right)\mu_b^{\;\;\dot b}({\bf P}_{\dot b})^{\gamma},\] and we obtain, similar to the previous discussion, that the branch points are in general connected to an infinite series of sheets, which can be reached by iterating (**P**\_a )\^\|\^n = \|W\_a\^b( **P**\_b )\^\|\^n-1 ,(**P**\_a )\^\^n = W\^b\_a( **P**\_b )\^\^n-1 , with \(W\), \(\bar{W}\) defined by \[W_a^{\;\;\dot b} = \left(\delta_a^b + {\bf P}_a {\bf P}^b\right)\mu_b^{\;\;\dot b}\,,\qquad\qquad \bar{W}_a^{\;\;\dot b} = \mu^{\dot c}_{\;\;a}\left(\delta_{\dot c}^{\dot b}-{\bf P}_{\dot c} {\bf P}^{\dot b}\right),\] with \(\dot{ W } = ( \bar{W} )^{-1}\), \(\dot{\bar{W}}= (W )^{-1}\) defined similarly by dotting/undotting all indices. As in the previous paragraph, we see that going around the branch point many times keeps leading to new sheets, since we expect in general that \((W \cdot \dot W )_a^{\;b}\neq \delta_a^b\), being there no reason to expect otherwise. We can summarise the finding of this section in a set of \({\bf P}\mu\) equations. For the first wing they read, \[({\bf P}_a)^{\bar{\gamma}} = {\bf P}_{\dot b} \mu^{\dot b }_{\;\; a}\,,\qquad\qquad ({\bf P}^a)^{\bar{\gamma}} = {\bf P}^{\dot b} \mu_{\dot b }^{\;\; a} ,\label{eq:Pmufirst}\] and \[(\mu_a^{\;\;\dot b})^{\gamma}-\mu_a^{\;\;\dot b} = {\bf P}_a ({\bf P}^{\dot b} )^{\bar{\gamma}}-({\bf P}_a)^\gamma {\bf P}^{\dot b}, \;\;\;\; (\mu^a_{\;\;\dot b})^{\gamma}-\mu^a_{\;\;\dot b} = -{\bf P}^a ({\bf P}_{\dot b} )^{\bar{\gamma}} + ({\bf P}^a)^\gamma {\bf P}_{\dot b}. \label{eq:Pmulast}\] Together with the mirror-periodicity of \(\mu\), this can also be taken as a self-consistent description of the QSC. As remarked for the \({\bf Q}\omega\)-system, these equations contain enough information to map the values of \({\bf P}\) and \(\mu\) functions on any sheet, back to the first main one.
# The ABA limit {#sec:ABAQSC}
In this section, we will find an asymptotic solution for some of the Q-functions in the large-\(L\) limit. This will lead us to a perfect match with the Asymptotic Bethe Ansatz for massive states, including the dressing phases.
## Large-volume scaling of the QSC
To deduce the large-\(L\) solution, we will use arguments developed for the \(AdS_5\) case in and then also successfully used for \(AdS_4\) case to derive the ABA in. The crucial observation is that, for large \(L\), some Q-functions are exponentially suppressed/enhanced, following the pattern of their large-\(u\) asymptotics ([\[eq:largeu\]](#eq:largeu){reference-type="ref" reference="eq:largeu"}). Following the notation of, we introduce a parameter \(\epsilon \propto e^{-L}\) to keep track of this scaling. We then see that for large \(L\) (i.e., \(\epsilon \sim 0\)), [\[eq:epsilonQ\]]{#eq:epsilonQ label="eq:epsilonQ"} &&Q\_a\|i \~(
), Q\^a\|i \~(
) ,\
&&**Q**\_i \~(\^-1, ), **Q**\^i \~(, \^-1 ), **P**\_i \~(, \^-1 ), **P**\^i \~(\^-1, ).[\[eq:epsilonQ2\]]{#eq:epsilonQ2 label="eq:epsilonQ2"} In the second wing, we would have exactly the same pattern for the dotted Q-functions. In addition, since the \(\omega\) functions are periodic on a Riemann section with short cuts, they have constant asymptotics. We will then assume that they all scale as \[\omega_k^{\;\dot m}\sim O(1) ,\; \omega^k_{\;m} \sim O(1),\; \omega_{\dot k}^{ \;m}\sim O(1), \; \omega^{\dot k}_{\;m}\sim O(1), \;\;\; \epsilon\rightarrow 0.\] We then notice that some of the QQ relations, \({\bf P}\mu\) and \({\bf Q}\omega\) equations simplify significantly. Dropping the subleading terms for \(\epsilon\rightarrow 0\) we find for instance, from ([\[eq:defmu\]](#eq:defmu){reference-type="ref" reference="eq:defmu"}), \[\mu_1^{\;\dot 2} = Q_{1|k}^-\omega^k_{\;\; \dot l} Q^{ \dot 2 | \dot l \;-} \sim Q_{1|1}^-\omega^1_{\;\; \dot 2} Q^{ \dot 2 | \dot 2 \;-} = Q_{1|1}^-\omega^1_{\;\; \dot 2} Q_{ \dot 1 | \dot 1}^{-},\] and similarly we get to \[\label{eq:muABAdef} \mu^{2}_{\;\;\dot 1} \sim Q_{1|1}^-\omega_2^{\;\; \dot 1} Q_{ \dot 1 | \dot 1}^{-},\] where we recalled that by definition \(Q^{2|2} = Q_{1|1}\). Another important equation is obtained starting from \({\bf P}_1 = (Q_{1|i}^+) {\bf Q}^i\), and considering the analytic continuation along \(\gamma\) (recall that \(Q_{a|1}^+\) has no cut on the real axis). Using the \({\bf Q}\omega\)-system, and then considering the large-\(L\) scaling, we get \[\label{eq:Ptilde} ({{\bf P}}_1 )^{\gamma} = Q_{1|l}^+ \omega^{l}_{\; \dot k} {\bf Q}^{\dot k} \sim Q_{1|1}^+ \omega^{1}_{\; \dot 2} {\bf Q}^{\dot 2},\] which will play an important role in the following derivation of the ABA. We now proceed to deduce the form of some of the elements of the QSC in the ABA scaling. To do that, we will take as a working hypothesis the property that, for the functions \(\mu_1^{\;\;\dot 2}\), \(\mu_{\dot 1}^{\;\;2}\), \(\mu^{\dot1}_{\;\;2}\), \(\mu^{1}_{\;\;\dot 2}\), the cut on the real axis becomes quadratic in the large-\(L\) limit. We will see that all the solutions for massive states fall into this category.[^10] We will also make an assumption that the gluing matrix follows the pattern one can deduce from the gluing equations in the classical limit, namely that all the diagonal elements vanish. Our derivation assumes that this is true at least in the ABA limit, but we suspect it may be true even at finite \(L\) (this is what happens in \(AdS_5\)). Finally, we will use the expressions obtained from ([\[eq:barmu\]](#eq:barmu){reference-type="ref" reference="eq:barmu"}) in the ABA limit, such as \[\label{eq:muABAbar} \mu_{1}^{\;\; \dot 2} \sim \bar{Q}_{1|1}^-( G_{\dot k}^{\;\; 1} \omega^{\dot k}_{\;\; l} G^l_{\;\;\dot 2} ) \bar{Q}_{\dot 1|\dot1}^-\propto \bar{Q}_{1|1}^-\omega^{\dot 2}_{\;\; 1} \bar{Q}_{\dot 1|\dot1}^-.\]
## Fixing Q-functions on the first sheet
#### Finding \(Q_{1|1}\), \(\mu_1^{\;\; \dot 2}\) and \(\omega^1_{\;\;\dot 2}\).
To determine these functions, we use the assumption on the quadratic nature of the branch point of \(\mu_1^{\;\;\dot 2}\) in the ABA limit. Even though this assumption could appear to be too restrictive, we will nevertheless show that in the ABA limit this extra restriction does not lead to any inconsistencies. The simplification of the analytic structure of \(\mu\) is quite typical in the ABA limit--for instance in the \(AdS_5\) case the discontinuity of \(\log\mu\) appears to be a simple rational function of \(x\), whereas in general it would have an infinity tower of cuts. With that in mind, we can follow closely, and this part may be skimmed through by the reader familiar with that paper. The surprises begin from section [4.3](#sec:surprises){reference-type="ref" reference="sec:surprises"}, where the non-quadratic nature of the branch points pops up again in a crucial way. We start by considering the function \(\mu_1^{\;\;\dot 2}(u+i/2)\). We take it to have a finite number of zeros on the first Riemann sheet with short cuts, and we store such zeros in a polynomial \(\mathbb{Q}(u)=\prod_i(u-u_i)\). We then consider \[(F)^2 \equiv \frac{\mu_1^{\;\; \dot 2} }{\mu_1^{\;\; \dot 2 ++} } \frac{\mathbb{Q}^+}{\mathbb{Q}^-} = \frac{\mu_1^{\;\; \dot 2} }{(\mu_1^{\;\; \dot 2})^{\gamma} } \frac{\mathbb{Q}^+}{\mathbb{Q}^-}\] which by definition has no zeros or poles on the first Riemann sheet with short cuts. Since by our assumption the branch points of \(\mu\) become quadratic in the ABA limit, using the property \((\mu)^{\gamma} = \mu^{++} \sim \mu^{\bar\gamma}\), it is simple to obtain the same equations as in: \[\label{eq:FRH} F (F)^{\gamma} \sim F (F)^{\bar\gamma} \sim \frac{\mathbb{Q}^+}{\mathbb{Q}^-}.\] All the other cuts in \(F\) must disappear in the ABA limit. In fact, using ([\[eq:muABAdef\]](#eq:muABAdef){reference-type="ref" reference="eq:muABAdef"}), and the periodicity of \(\omega\), we see that \(F^2\) can be rewritten as \[\label{eq:Q11mu} F^2 =\frac{Q_{1|1}^-\, Q_{\dot 1|\dot 1}^-}{Q_{1|1}^+ \, Q_{\dot 1|\dot 1}^+}\frac{\mathbb{Q}^+}{\mathbb{Q}^-},\] which does not have cuts in the upper half plane, while ([\[eq:muABAbar\]](#eq:muABAbar){reference-type="ref" reference="eq:muABAbar"}) leads us to the expression \[F^2 = \frac{\bar{Q}_{1|1}^-\, \bar{Q}_{\dot 1|\dot 1}^-}{\bar{Q}_{1|1}^+ \, \bar{Q}_{\dot 1|\dot 1}^+}\frac{\mathbb{Q}^+}{\mathbb{Q}^-},\] which shows that there are no cuts in the lower half plane either. Taking into account that \(F\) has constant asymptotics at large \(u\) on the first sheet, we have a simple Riemann-Hilbert problem ([\[eq:FRH\]](#eq:FRH){reference-type="ref" reference="eq:FRH"}), with the standard solution \[F = \pm e^{i \frac{\mathcal{P}}{2} } \frac{B_{(+)}}{B_{(-)}},\] with \(e^{i\mathcal{P}}\equiv\prod_{i} \frac{x_i^+}{x_i^-}\), and \(B_{(\pm)}(u)\equiv \sqrt{\frac{h}{x_i^{\mp}}} (\frac{1}{x(u)}-x_i^{\mp} )\). The constant factor will not be very important in the current considerations.[^11] Setting \(Q_{1|1} Q_{\dot 1|\dot 1} \equiv \mathbb{Q} (f^+)^2\), equation ([\[eq:Q11mu\]](#eq:Q11mu){reference-type="ref" reference="eq:Q11mu"}) then gives us a difference equation \[\frac{f^{++}}{f} = \frac{B_{(-)}}{B_{(+)}},\] where by construction \(f\) should have neither poles nor zeros in the upper half plane, and power-like asymptotics. Up to a multiplicative constant, the solution is \[\label{eq:intf} f(u) \propto \text{exp} \left(\int_{-2h}^{2h}\frac{dz}{2 \pi i}\log\frac{B_{(-)}(z + i 0^+)R_{(+)}(z+ i 0^+)}{B_{(+)}(z+ i 0^+)R_{(-)}(z+ i 0^+)} \partial_z \log\Gamma( i z-iu ) \right),\] where we use \(\propto\) to indicate that there could be an irrelevant constant factor in the equation. With the explicit form of \(f\) in , we have fixed \(Q_{1|1} Q_{\dot 1|\dot 1}\) completely. Noticing that \(\mu_1^{\;\;\dot 2} = Q_{1|1}^-Q_{\dot 1|\dot 1}^-\omega^1_{\;\;\dot 2} \propto \bar{Q}_{1|1}^-\bar{Q}_{\dot 1|\dot 1}^-\omega^1_{\;\;\dot 2}\), where \(\omega\) should be \(i\)-periodic, we can also find \[\mu_1^{\;\;\dot 2} \propto\mathbb{Q}^-f \bar{f}^{--}, \;\;\; \omega^1_{\;\;\dot 2} \propto \frac{\bar{f}^{--} }{f}, \;\;\; Q_{1|1} Q_{\dot 1| \dot 1} \propto \mathbb{Q} \, ( f^+ )^2,\] where \(\bar{f}\) is solution of \(\bar{f}/\bar{f}^{--} = \frac{B_{(-)}}{B_{(+)}}\) with no cuts in the lower half plane and constant asymptotics.[^12] From the expression ([\[eq:Q11mu\]](#eq:Q11mu){reference-type="ref" reference="eq:Q11mu"}), we also see that the set of zeros of \(\mathbb{Q}\) must coincide with the union of the zeros of \(Q_{1|1}\) and \(Q_{\dot 1|\dot 1}\). Therefore we split this polynomial as \(\mathbb{Q}(u) \equiv \mathbb{Q}_2(u) \mathbb{Q}_{\dot 2}(u)\), with \(\mathbb{Q}_2(u)\equiv \prod_{i=1}^{K_2}(u-u_{2,i})\), \(\mathbb{Q}_{\dot 2}(u)\equiv \prod_{i=1}^{K_{\dot 2}}(u-u_{\dot 2,i})\), with the understanding that \(\mathbb{Q}_2\) contains zeros of \(Q_{1|1}\), and \(\mathbb{Q}_{\dot 2}\) zeros of \(Q_{\dot1|\dot1}\). This notation is chosen in anticipation of the role of the zeros in the ABA. With these conventions, we have \[\label{eq:expl1} \mu_1^{\;\;\dot 2} \propto \mathbb{Q}^-_2 \mathbb{Q}^-_{\dot 2} f_2 f_{\dot 2} \bar{f}^{--}_2 \bar{f}^{--}_{\dot 2}, \;\;\; \omega^1_{\;\;\dot 2 } \propto \frac{\bar{f}^{--}_2 \bar{f}^{--}_{\dot 2} }{f_2 \, f_{\dot 2}}, \;\;\; Q_{1|1} \propto \mathbb{Q}_2 \, f^+_2 f^+_{\dot 2}\, P, \;\;\; Q_{\dot 1|\dot 1} \propto \mathbb{Q}_{\dot 2} \, f^+_2 f^+_{\dot 2} \frac{1}{P},\] with the obvious notation that \(f_{\alpha}\) are solutions of \(f_{\alpha}^{++}/f_{\alpha}= \frac{B_{\alpha,(-)}}{B_{\alpha,(+)}}\), with \(\alpha \in \left\{2,\dot 2\right\}\) (see appendix [\[app:ABA\]](#app:ABA){reference-type="ref" reference="app:ABA"}), and where \(P\) is a yet unfixed function of \(u\) coming from splitting the product \(Q_{1|1} Q_{\dot 1|\dot 1}\). This function should have neither zeros nor poles, and moreover \(P^-\) cannot have any cuts in the upper half plane. On the other hand, the quantity \[\label{eq:LHPA4} \bar{Q}_{1|1}^+ = Q_{1|j}^+ \Omega_1^{\;\;j} \sim Q_{1|1}^+ \Omega_1^{\;\;1} \;\;\;\text{ for } \epsilon\rightarrow 0,\] should be analytic in the lower half plane, where the matrix \(\Omega\) is defined by \(\omega_{\dot a}^{\;\; b} = G_{\dot a}^{\;\; c} \, \Omega_{ c}^{\;\; b}\). Using the assumed classics-inspired off-diagonal property of the gluing matrix, we see that \(\Omega_1^{\;\;1} \propto \omega_{\dot 2}^{\;\; 1}\). Then, from ([\[eq:LHPA4\]](#eq:LHPA4){reference-type="ref" reference="eq:LHPA4"}) and the above found solution for \(\omega_{\dot 2}^{\;\; 1}\), we deduce that \(\bar{Q}_{1|1}^+\)-which should be analytic in the lower half plane-can also be written as \(\mathbb{Q}_2 \frac{B_{2,(-)} B_{\dot 2,(-)}}{B_{2,(+)} B_{\dot 2,(+)}} \bar{f}_2^{--}\, \bar{f}_{\dot 2}^{--}\, P^+.\) Since all the other factors already have this property, we conclude that \(P^+\) should have no cuts in the lower half plane. All together, we found that the function \(P\) cannot have any singularities or zeroes and thus is a constant (due to regularity at infinity). In conclusion, we found && Q\_1\|1 \_2 f\_\^+ f_2\^+, Q\_\| \_ f\_\^+ f_2\^+ ,[\[eq:Q11fixed\]]{#eq:Q11fixed label="eq:Q11fixed"}\
&&\_1\^ \_\^ \_2\^- \_\^-f_2 \|f\_ 2\^-- f\_ \|f\_ \^--,\
&& \^1\_ \^\_ ,[\[eq:omegas\]]{#eq:omegas label="eq:omegas"} where we included the values of more \(\omega\), \(\mu\) functions, obtained by obvious generalisations of the argument above.
#### Parametrising \({\bf P}\) and \({\bf Q}\) functions.
From the \(AdS_5\) and \(AdS_4\) cases, we expect that a special subset of \({\bf P}\) and \({\bf Q}\) functions will converge to simple explicit expressions in the ABA limit. This is the subset of the \({\bf P}\) functions which are *small*, together with the \({\bf Q}\) functions that are *large*, for \(\epsilon\rightarrow 0\). From ([\[eq:epsilonQ2\]](#eq:epsilonQ2){reference-type="ref" reference="eq:epsilonQ2"}), we see that those are \({\bf P}_1\), \({\bf P}^2\), \({\bf Q}_1\), \({\bf Q}^2\), and their dotted counterparts. We expect that their zeros on the first sheet will acquire the meaning of Bethe roots. With this in mind, we make the following ansatz: \[\begin{aligned}
\label{eq:Qide} {\bf P}_1 &\propto x^{-L/2} \,\mathcal{A} \times R_{\tilde{1} } B_{\tilde{\dot 1}} \, B_{2, (-)} , & {\bf P}^2 &\propto x^{-L/2} \,\mathcal{A} \times R_{\tilde 3} B_{\tilde{\dot 3}} B_{2, (-) } ,\\ {\bf Q}_1 &\propto \frac{x^{L/2} }{\mathcal{A}' } \times R_{1} B_{\dot 1} f_{2}\frac{f_{\dot 2} }{B_{\dot 2, (+) } }, & {\bf Q}^2 &\propto \frac{x^{{L}/2} }{\mathcal{A}'} \times R_{3} B_{\dot 3} \, f_{2} \frac{f_{\dot 2} }{B_{\dot 2, (+) } }\;.\label{eq:Qidelast} \end{aligned}\] Above, we have stored the zeros of the \({\bf P}\) and \({\bf Q}\) functions on the first sheet inside the Zhukovsky polynomials \(R_{\alpha}\), defined[^13] in appendix [\[app:ABA\]](#app:ABA){reference-type="ref" reference="app:ABA"} (again, the notation anticipates the role of these zeros in the ABA, but for now they are generic parameters). The other \(B_{\alpha}\) and \(f_{\alpha}\) factors (also defined in the appendix) are chosen for future convenience, but they do not have zeros on the first sheet. Notice that the ansatz above is fully general, because it contains the arbitrary functions of \(u\) \(\mathcal{A}(u)\), \(\mathcal{A}'(u)\). By construction they should have no poles or zeros on the first sheet, and moreover \(\mathcal{A}\), which appears in the \({\bf P}\) functions, can have only a single cut. Comparing with , we see that we can write the important function \(r\) in two alternative ways as \[r\propto \frac{R_{\tilde 1}B_{\tilde {\dot 1}}}{R_{\tilde 3}B_{\tilde{\dot 3}}}\propto \frac{R_3 B_{\dot 3}}{R_1 B_{\dot 1}}\;,\] which means that \(R_{\tilde 3}\) and \(R_{\tilde 1}\) could have common zeroes. Likewise we make a similar ansatz for the second wing: \[\begin{aligned}
{\bf P}_{\dot 1} &\propto x^{-{L}/2} \,\dot{\mathcal{A}} \times R_{\dot 3}' B_{3}' \, B_{\dot 2,(-)}, & {\bf P}^{\dot 2} &\propto x^{-{L}/2} \,\dot{\mathcal{A}} \times R_{\dot 1}' B_{1}' \, B_{\dot 2, (-) }, \label{eq:Qsnd}\\ {\bf Q}_{\dot 1} &\propto \frac{x^{{L}/2} }{\dot{\mathcal{A}}' } \times R_{\tilde{\dot 3} }' B_{\tilde 3 }' f_{\dot 2} \, \frac{f_{2} }{B_{2, (+) } }, & {\bf Q}^{\dot 2} &\propto \frac{x^{{L}/2} }{\dot{\mathcal{A}}' } \times R_{\tilde{\dot 1}}' B_{\tilde 1}' \, f_{\dot 2} \, \frac{f_{2} }{B_{2,(+) } }\;,\label{eq:Qsndlast} \end{aligned}\] with functions \(\dot{\mathcal{A}}(u)\), \(\dot{\mathcal{A}}'(u)\) having no zeroes on the main sheet. In ([\[eq:Qsnd\]](#eq:Qsnd){reference-type="ref" reference="eq:Qsnd"}),([\[eq:Qsndlast\]](#eq:Qsndlast){reference-type="ref" reference="eq:Qsndlast"}), we have introduced polynomials in \(x\) vs \(\frac{1}{x}\), \(R_{\alpha}'\) and \(B_{\alpha}'\), respectively. They are defined just like in ([\[eq:defB\]](#eq:defB){reference-type="ref" reference="eq:defB"}),([\[eq:defR\]](#eq:defR){reference-type="ref" reference="eq:defR"}), but where the zeros of these polynomials (and their number) are in principle unrelated to the ones appearing in the first wing. We will however soon see that there is a simple identification. From we again get \[\dot r \propto \frac{R'_{\dot 3} B'_3}{R'_{\dot 1} B'_1}\propto \frac{R'_{\tilde{\dot 1}} B'_{\tilde 1}}{R'_{\tilde{\dot 3}} B'_{\tilde 3}}\;.\] Furthermore, recalling that \(\dot r^\gamma=r\) we get \[\la{RBall} \frac{R_{\tilde 1}B_{\tilde {\dot 1}}}{R_{\tilde 3}B_{\tilde{\dot 3}}}\propto \frac{R_3 B_{\dot 3}}{R_1 B_{\dot 1}} \propto \frac{B'_{\dot 3} R'_3}{B'_{\dot 1} R'_1}\propto \frac{B'_{\tilde{\dot 1}} R'_{\tilde 1}}{B'_{\tilde{\dot 3}} R'_{\tilde 3}}\;.\] One can for example deduce that \(R_{\tilde 1} R_1=R_3R_{\tilde 3}\) etc. from the above equation.
#### Fermionic duality equation.
An important constraint comes from one of the QQ relations \[\label{eq:QQtwoeq} Q_{1|1}^+-Q_{1|1}^-= {\bf Q}_1 {\bf P}_1\;,\] where we see the appearance of \(Q_{1|1}\) determined in ([\[eq:Q11fixed\]](#eq:Q11fixed){reference-type="ref" reference="eq:Q11fixed"}). Plugging in that value, and the ansatz ([\[eq:Qide\]](#eq:Qide){reference-type="ref" reference="eq:Qide"}),([\[eq:Qidelast\]](#eq:Qidelast){reference-type="ref" reference="eq:Qidelast"}), we find, from the first equality in ([\[eq:QQtwoeq\]](#eq:QQtwoeq){reference-type="ref" reference="eq:QQtwoeq"}), \[\la{RBall2} R_{2,(+)} B_{\dot 2, (-)}-R_{2, (-)} B_{\dot 2, (+)} \propto R_1 \, R_{\tilde 1} B_{\dot 1} B_{\tilde{ \dot 1}} \frac{\mathcal{A}}{\mathcal{A}'}\;,\] where we used the property that \(\mathbb{Q}_{\alpha}^{\pm} =B_{\alpha, (\pm)}R_{\alpha, (\pm)}\). Since the left hand side is a rational function in \(1/x(u)\), and \(\mathcal{A}\), \(\mathcal{A}'\) should have no zeros or poles on the first sheet, the ratio \(\mathcal{A}(u)/\mathcal{A}'(u)\) can only be a polynomial in the variable \(\frac{1}{x(u)}\). But we can absorb any such function in a redefinition of the \(B_{\tilde{\dot 1}}\), \(B_{\dot 1}\) polynomials (which are so far completely unconstrained), so without loss of generality we can take \(\mathcal{A}/\mathcal{A}' = 1\). Similar considerations arise from considering ([\[eq:QQtwoeq\]](#eq:QQtwoeq){reference-type="ref" reference="eq:QQtwoeq"}) in the second wing. From now on, therefore we take \[\mathcal{A}(u) = \mathcal{A}'(u) \;\;,\;\; \dot{\mathcal{A}}(u) = {\dot{\mathcal{A}}}'(u).\] Notice that we still have two undetermined functions, which will be fixed in the next section. Using that \({\cal A}={\cal A}'\), from ([\[RBall2\]](#RBall2){reference-type="ref" reference="RBall2"}) we obtain \[\label{eq:fermdual1} R_{2,(+)} B_{\dot 2, (-)}-R_{2, (-)} B_{\dot 2, (+)} \propto R_1 \, R_{\tilde 1} B_{\dot 1} B_{\tilde{ \dot 1}}\;.\] The analogous constraint obtained by considering the second wing reads \[\label{eq:fermdual2} B_{2, (-)} R_{\dot 2, (+)}-B_{2,(+)} R_{\dot 2, (+)} \propto B_1' \, B_{\tilde 1}' R_{\dot 1}' R_{\tilde{ \dot 1}}'\;,\] and analytically continuing this equation to another sheet we find the identity \[R_1 \, R_{\tilde 1} B_{\dot 1} B_{\tilde{ \dot 1}} \propto R_1' \, R_{\tilde 1}' B_{\dot 1}' B_{\tilde{ \dot 1}}'\] which implies \[R_1 \, R_{\tilde 1}=R_1' \, R_{\tilde 1}'\;.\] Equations of the form ([\[eq:fermdual1\]](#eq:fermdual1){reference-type="ref" reference="eq:fermdual1"}) are examples of fermionic duality relations. They imply that the sets of roots with labels \(1\), \(2\), \(3\) (or alternatively the "dual" set obtained with \(1\leftrightarrow \tilde{1}\), \(3\leftrightarrow \tilde{3}\)) satisfy the auxiliary ABA equations of the form \[\label{eq:BAauxiliary} 1 =\left. \frac{\mathbb{Q}^+_2 B_{2,(-)}\,B_{\dot 2,(-)}}{\mathbb{Q}^-_2 B_{2,(+)}\,B_{\dot 2,(+)}} \right|_{u \in \left\{\text{roots of type } 1,\tilde{1}, 3,\tilde{3}\right\}}.\]
## Going inside the cut: fixing the dressing phases {#sec:surprises}
So far we reduced the ansatz for \({\bf P}\)'s and \({\bf Q}\)'s to just two unknown functions with one cut and no zeroes \({\cal A}\) and \(\dot{\cal A}\) on the main sheet. In order to constrain them further, we need to go to the next sheet of their Riemann surfaces. This will bring us to the most interesting part of the analysis, where things will be radically different than in \(AdS_5\) and \(AdS_4\). By studying equations of the form ([\[eq:Ptilde\]](#eq:Ptilde){reference-type="ref" reference="eq:Ptilde"}), which we repeat here, [\[eq:P1tildeABA2\]]{#eq:P1tildeABA2 label="eq:P1tildeABA2"} && (**P**\_1)\^ \~Q\_1\| 1\^+ \^ 1\_ **Q**\^ , (**P**\_)\^ \~Q\_\| \^+ \^ \_ **Q**\^ 2 [\[eq:P1tildeABA3\]]{#eq:P1tildeABA3 label="eq:P1tildeABA3"} we will find that the \({\bf P}\) and \({\bf Q}\) functions *cannot have* a quadratic cut even in the ABA limit. We will also be able to fix the form of the yet undetermined functions \(\mathcal{A}(u)\), \(\dot{\mathcal{A}}(u)\) and relate them to the dressing phases of .
### The cuts cannot be quadratic {#sec:notquadratic}
The strategy will be to compare the r.h.s. of each of the equations ([\[eq:P1tildeABA2\]](#eq:P1tildeABA2){reference-type="ref" reference="eq:P1tildeABA2"}), with the analytic continuation of \({\bf P}\) functions, starting from their form in ([\[eq:Qide\]](#eq:Qide){reference-type="ref" reference="eq:Qide"}),([\[eq:Qsndlast\]](#eq:Qsndlast){reference-type="ref" reference="eq:Qsndlast"}).[^14] From the first equation in ([\[eq:Ptilde\]](#eq:Ptilde){reference-type="ref" reference="eq:Ptilde"}), in particular, we obtain: \[{\bf P}_1^\gamma=x^{L/2}({\mathcal{A}} )^{\gamma} B_{\tilde{1} } R_{\tilde{\dot 1}} \, R_{2, (-)}= \({\mathbb Q}_2^+ f_{\dot 2}^{++} f_{2}^{++} \) \( \frac{\bar f_2\bar f_{\dot 2}}{ f_2^{++} f_{\dot 2}^{++} } \) \( \frac{x^{L/2}}{\dot{\cal A}} R_{\tilde {\dot 1}}' B'_{\tilde 1} \frac{f_2 f_{\dot 2}}{B_{2,(+)}} \)\;.\] We noticed in the previous section that the roots of \(R_{\tilde 1}\) and \(R'_{\tilde 1}\) satisfy the same BAE equation . The same is true for the roots of \(R_{\tilde {\dot 1}}\) and \(R'_{\tilde {\dot 1}}\). Whereas this does not necessarily mean that all roots coincide, we will assume \(R_{\tilde 1}=R'_{\tilde 1}\) and \(R_{\tilde {\dot 1}}=R'_{\tilde {\dot 1}}\). In this case we get a nice cancellation in the above equation, which further supports this requirement. Then we get a simple relation \[\label{eq:AtildeD} ({\mathcal{A}} )^{\gamma} \, \dot{\mathcal{A}} = \left(\frac{ R_{2,(+)} }{R_{2,(-)} } \right) \left( \bar{f}_2^{--} f^{++}_2 \bar{ f}_{\dot 2}^{--} {f}_{\dot 2}^{++}\right)\;.\] It is striking to compare this with the consequence of the second relation in ([\[eq:P1tildeABA3\]](#eq:P1tildeABA3){reference-type="ref" reference="eq:P1tildeABA3"}), which yields \[\label{eq:DtildeA} {\mathcal{A}} \, (\dot{\mathcal{A}})^{\gamma} = \left(\frac{ R_{\dot 2, (+)} }{R_{\dot 2, (-)} } \right) \, \left( \bar{f}_2^{--} f^{++}_2 \bar{ f}_{\dot 2}^{--} {f}_{\dot 2}^{++}\right).\] Now we continue this relation along the reverse path \(\bar\gamma\): the result on the l.h.s. is \((\mathcal{A})^{\bar\gamma}\dot{\mathcal{A}}\), and the analytic continuation of the r.h.s. is simple to compute, since the \(f^{++}\), \(\bar f^{--}\) functions have no cut on the real axis, so are left unchanged. By comparing the result with ([\[eq:AtildeD\]](#eq:AtildeD){reference-type="ref" reference="eq:AtildeD"}), we get the following "double-discontinuity\" relations \[\la{doubledisc} \frac{\mathcal{A} ^{\gamma}}{ \mathcal{A} ^{\bar\gamma}} = \frac{ R_{2,(+)} }{R_{2,(-)} } \frac{B_{\dot 2, (-)} }{ B_{\dot 2, (+)} }\,,\qquad\qquad \frac{\mathcal{\dot A} ^{\gamma}}{ \mathcal{\dot A} ^{\bar\gamma}} = \frac{ R_{\dot 2,(+)} }{R_{\dot 2,(-)} } \frac{B_{2, (-)} }{ B_{2, (+)} }\;,\] where the r.h.s. clearly cannot vanish (except for the vacuum) since the \(R_{\alpha}\) and \(B_{\alpha}\) functions have zeros on different sheets. We will now solve and .
### Relation to the dressing phases {#sec:dressingrel}
In order to find the solution, without lack of generality we introduce the following ansatz in terms of \(\rho\) and \(\dot \rho\) = \^1, \_2 \^1,\_ , = \^1,\_2 \^1,\_ , where, using notation from, \(\sigma^{1,{\rm BES}}_\alpha\) denote natural building blocks of the Beisert-Eden-Staudacher dressing factor. They satisfy \[( \sigma^{1,{\rm BES}}_{\alpha} )^{\gamma} \; \sigma^{1,{\rm BES}}_{\alpha} = ( \sigma^{1,{\rm BES}}_{\alpha} )^{\bar\gamma} \; \sigma^{1,{\rm BES}}_{\alpha} \propto f^{ ++}_{\alpha} \bar{f}^{--}_{\alpha \;\;,\;\; \alpha=2,\dot2\;,\] and are related to the product of the BES dressing factors via \[\sigma_{\rm BES}(u)=\frac{\sigma^{1,\text{BES}}(u+\tfrac{i}{2}) }{\sigma^{1,\text{BES}}(u-\tfrac{i}{2}) }\;,\] with the notation explained in appendix [\[app:ABA\]](#app:ABA){reference-type="ref" reference="app:ABA"}. With this redefinition, , ([\[eq:DtildeA\]](#eq:DtildeA){reference-type="ref" reference="eq:DtildeA"}) become \[\label{eq:lastRH} ( \rho )^{\gamma} \dot\rho \propto \sqrt{\frac{R_{2,(+)}}{R_{2,(-)}}\; \frac{B_{\dot 2,(-)}}{B_{\dot 2, (+)}} }\,,\qquad\qquad (\dot\rho )^{\gamma} \,\rho \propto \sqrt{\frac{R_{\dot 2,(+)}}{R_{\dot 2,(-)}}\; \frac{B_{ 2,(-)}}{B_{ 2, (+)}} }\;.\] In appendix [\[app:crossing\]](#app:crossing){reference-type="ref" reference="app:crossing"}, we define the functions \(\sigma^{1,{\rm extra}}\) and \(\tilde\sigma^{1,{\rm extra}}\) which are related to the two independent dressing phases appearing in ABA equations of section [2.1](#sec:ABA){reference-type="ref" reference="sec:ABA"} in the following way: \[\sigma(u) = \sigma_{\rm BES}(u) \frac{\sigma^{1,{\rm extra}}(u+\tfrac i2)}{\sigma^{1,{\rm extra}}(u-\tfrac i2)}\,,\qquad\qquad \tilde{\sigma}(u) = \sigma_{\rm BES}(u) \frac{\tilde{\sigma}^{1,{\rm extra}}(u+\tfrac i2)}{\tilde{\sigma}^{1,{\rm extra}}(u-\tfrac i2)}.\] In the same appendix, we also show that these extra pieces satisfy the following identities [\[eq:crossstrange\]]{#eq:crossstrange label="eq:crossstrange"} ( \^1,extra\_ )\^ \_\^1,extra = , ( \^1,extra\_ )\^ \_\^1,extra= ,{2,} , which we both verify directly and also independently deduce from crossing via functional arguments. Using those building blocks, we can write \[\rho = {\sigma}_2^{1,{\rm extra}}{\tilde \sigma}_{\dot 2}^{1,{\rm extra}}\rho_0\;\;,\;\; \dot \rho = {\sigma}_{\dot 2}^{1,{\rm extra}}{\tilde \sigma}_{2}^{1,{\rm extra}}\dot \rho_0\;,\] where \(\rho_0\) and \(\dot\rho_0\) should be functions with square-root branch cut on the real axis satisfying \[\rho^\gamma_0 = 1/\dot\rho_0\;\;,\;\; \dot \rho^\gamma_0 = 1/\rho_0\;.\] This equation tells us that \(\rho^\gamma_0\) is a function with a single cut and neither zeroes nor poles, and likewise \(\dot\rho_0\) and \(\rho_0\). In other words it can only be a power of \(x\), which can be included into a re-definition of \(L\). So without reducing the generality we can set \(\dot\rho_0=\rho_0=1\). This completes the derivation of the asymptotic limit of our QSC.
## Summary of results for the asymptotic limit
Let us summarise what we found for the expressions of \({\bf P}\) and \({\bf Q}\) functions. In the first wing we have: \[\begin{aligned}
{\bf P}_1 &\propto x^{-L/2} \, R_{\tilde{1} } B_{\tilde{\dot 1}} \, \sqrt{B_{2, (+) } B_{2, (-) } }\;\sigma^{1}_2 \, \tilde{\sigma}^{1}_{\dot 2}, & {\bf P}^2 &\propto x^{-L/2} \, R_{\tilde 3} B_{\tilde{\dot 3}} \sqrt{B_{2, (+) } B_{2, (-) } }\;\sigma^{1}_2 \, \tilde{\sigma}^{1}_{\dot 2},\label{eq:Qfirstwingfinal0}\\ {\bf Q}_1 &\propto x^{L/2 } R_{1} B_{\dot 1} \,\sqrt{\frac{B_{2,(-)}}{B_{2,(+)}}}\; \frac{f_2\,f_{\dot 2} }{ B_{\dot 2, (+) } \, \sigma^{1}_2 \, \tilde{\sigma}^{1}_{\dot 2}}, & {\bf Q}^2 & \propto x^{{L}/2} R_{3} B_{\dot 3} \,\sqrt{\frac{B_{2,(-)}}{B_{2,(+)}}} \; \frac{f_2\,f_{\dot 2} }{ B_{\dot 2, (+) } \, \sigma^{1}_2 \, \tilde{\sigma}^{1}_{\dot 2}},\label{eq:Qfirstwingfinal} \end{aligned}\] and in the second wing: \[\begin{aligned}
{\bf P}_{\dot 1} &\propto x^{-{L}/2} \, R_{\dot 3} B_{3} \, \sqrt{B_{\dot 2,(+)} B_{\dot 2,(-)}}\; \;\sigma^{1}_{\dot 2} \, \tilde{\sigma}^{1}_{ 2}, & {\bf P}^{\dot 2} & \propto x^{-{L}/2} \, R_{\dot 1} B_{1} \,\sqrt{B_{\dot 2,(+)} B_{\dot 2,(-)}}\; \;\sigma^{1}_{\dot 2} \, \tilde{\sigma}^{1}_{ 2}, \label{eq:Qsndwingfinal0} \\ {\bf Q}_{\dot 1} &\propto {x^{{L}/2} }\, R_{\tilde{\dot 3} } B_{\tilde 3 } \,\sqrt{\frac{ B_{\dot 2, (-)} }{ B_{\dot 2, (+)} }} \, \frac{f_{2} f_{\dot 2} }{B_{2, (+) } \, \sigma^{1}_{\dot 2} \, \tilde{\sigma}^{1}_{ 2} }, &{\bf Q}^{\dot 2} &\propto {x^{{L}/2} }\, R_{\tilde{\dot 1}} B_{\tilde 1} \,\sqrt{\frac{ B_{\dot 2, (-)} }{ B_{\dot 2, (+)} }} \, \frac{f_{2} f_{\dot 2} }{B_{2, (+) } \, \sigma^{1}_{\dot 2} \, \tilde{\sigma}^{1}_{2} }.\label{eq:Qsndwingfinal} \end{aligned}\] All relevant notations are collected in Appendix [\[app:ABA\]](#app:ABA){reference-type="ref" reference="app:ABA"}. Having an asymptotic solution for all relevant \({\bf P}\) and \({\bf Q}\) functions we can plug them into the exact Bethe ansatz equations - and compare the result with the ABA .
## Match with the Asymptotic Bethe Ansatz
We have finally arrived at a full specification of the Q-functions \({\bf P}_1\), \({\bf P}^2\), \({\bf Q}_1\), \({\bf Q}^2\), \(Q_{1|1}\), and their dotted cousins, in the ABA limit. To obtain the Asymptotic Bethe Ansatz, we can just plug their values in the exact Bethe equations - following from the Q-system. We will have the following correspondence between the zeros appearing on the first sheet of the Q-functions, and the Bethe roots appearing in the Asymptotic Bethe Ansatz: for the first wing \[\begin{array}{cccc} \text{Roots:} & u_{1,k} & u_{2,k} & u_{3,k} \\ \text{Q-function:} & {\bf Q}_{1} & Q_{1|1} & {\bf Q}^2 \end{array} ,\;\;\;\;\; \begin{array}{ccc} \text{Dual roots:} & {u}_{\tilde{1},k} &{u}_{\tilde{3},k} \\ \text{Q-function:} & {\bf P}_{1} & {\bf P}^2 \end{array},\] and for the second wing: \[\begin{array}{cccc} \text{Root:} & u_{\dot 1,k} & u_{\dot2, k}& u_{\dot 3, k} \\ \text{Q-function:} & {{\bf P}}_{ \dot 1} & {Q}_{\dot 1|\dot 1} & {\bf P}^{\dot 2} \end{array}, \;\;\;\;\; \begin{array}{ccc} \text{Dual roots:} & {u}_{\tilde{\dot 1},k} &{u}_{\tilde{\dot 3},k} \\ \text{Q-function:} & {\bf Q}_{\dot 1} & {\bf Q}^{\dot 2 } \end{array}.\] In particular, the exact Bethe equations ([\[eq:ABA1L\]](#eq:ABA1L){reference-type="ref" reference="eq:ABA1L"})-([\[eq:ABA3L\]](#eq:ABA3L){reference-type="ref" reference="eq:ABA3L"}) for the first wing reduce precisely the ABA equations ([\[eq:Bethe13\]](#eq:Bethe13){reference-type="ref" reference="eq:Bethe13"})-([\[eq:Bethe33\]](#eq:Bethe33){reference-type="ref" reference="eq:Bethe33"}) using the Q-functions ([\[eq:Qfirstwingfinal0\]](#eq:Qfirstwingfinal0){reference-type="ref" reference="eq:Qfirstwingfinal0"})-([\[eq:Qfirstwingfinal\]](#eq:Qfirstwingfinal){reference-type="ref" reference="eq:Qfirstwingfinal"}). Similarly, using the exact Bethe equations of the form ([\[eq:ABA1Lb\]](#eq:ABA1Lb){reference-type="ref" reference="eq:ABA1Lb"})-([\[eq:ABA3Lb\]](#eq:ABA3Lb){reference-type="ref" reference="eq:ABA3Lb"}), but for the dotted Q-functions, we reproduce the ABA equations ([\[eq:Bethe21\]](#eq:Bethe21){reference-type="ref" reference="eq:Bethe21"})-([\[eq:Bethe23\]](#eq:Bethe23){reference-type="ref" reference="eq:Bethe23"}) using the asymptotic values ([\[eq:Qsndwingfinal0\]](#eq:Qsndwingfinal0){reference-type="ref" reference="eq:Qsndwingfinal0"})-([\[eq:Qsndwingfinal\]](#eq:Qsndwingfinal){reference-type="ref" reference="eq:Qsndwingfinal"}). As an example to demonstrate the procedure, we display the case of the middle-node equation for the first wing. At the roots of \(Q_{1|1}\) i.e. at \(u=u_{2,i}\) we have \[\begin{aligned}
&&-1 = \frac{Q_{1|1}^{++} Q_{\emptyset|1}^-Q_{12|1}^-}{Q_{1|1}^{--} Q_{\emptyset|1}^+ Q_{12|1}^+} = \frac{Q_{1|1}^{++} {\bf Q}_{1}^-{\bf Q}^{2-}}{Q_{1|1}^{--} {\bf Q}_{1}^+ {\bf Q}^{2+}} \nonumber\\ &&\qquad \qquad = \frac{(x^-)^L (\sigma_2^{1+}\tilde{\sigma}_{\dot{2}}^{1+})^2 {\mathbb Q}_2^{++} R_1^-B_{\dot{1}}^- R_3^-B_{\dot{3}}^- f_2^{[+3]}f_{\dot{2}}^{[+3]}f_2^-f_{\dot{2}}^- B_{2(-)}^- B_{2(+)}^+ [B_{\dot{2}(+)}^+]^2} { (x^+)^L (\sigma_2^{1-}\tilde{\sigma}_{\dot{2}}^{1-})^2 {\mathbb Q}_2^{--} R_1^+ B_{\dot{1}}^+R_3^+ B_{\dot{3}}^+ f_2^+\;\;\; f_{\dot{2}}^+\;\;\;f_2^+f_{\dot{2}}^+B_{2(-)}^+B_{2(+)}^-[B_{\dot{2}(+)}^-]^2 },\nonumber \end{aligned}\] where we have cancelled some terms repeated in the numerator and denominator. Next we have to use the defining property of the function \(f_\alpha\): \(\frac{f_\alpha^{++}}{f_\alpha} = \frac{B_{\alpha,(-)}}{B_{\alpha,(+)}}\) in appropriate shifted version, to re-create various \(B\) functions, some of which then cancel out and some remain. At the end of this massive simplification what is left is exactly the middle-node ABA equation for the first wing, where one needs to recall how the dressing phases are reconstructed from \(\sigma^1\) and \(\tilde{\sigma}^1\) via \(\sigma_\alpha =\frac{\sigma^{1+}}{\sigma^{1-}}\) and \(\tilde\sigma_\alpha =\frac{\tilde\sigma^{1+}_{\alpha}}{\tilde\sigma^{1-}_{\alpha}}\): \[\begin{aligned}
-1 &=& \left(\frac{x^{-} }{x^{+}}\right)^{L} \times \frac{\mathbb{Q}^{++}_{2}}{\mathbb{Q}^{--}_{2} }\times \left(\sigma_{2} \right)^2 \times \frac{R_{1}^-R_{3}^-}{ R_{1}^+ R_{3}^+ }\nonumber\\ &&\times \left. \frac{B_{\dot 2,(-) }^+ \, B_{\dot 2, (+) }^+}{B_{\dot 2, (-) }^-\, B_{\dot 2, (+) }^-}\times \left( {\tilde\sigma}_{\dot 2} \right)^2\times \frac{B_{\dot 1}^-B_{\dot 3}^-}{ B_{\dot 1}^+ B_{\dot 3}^+ } \right|_{u = u_{2,i}},\,\; i=1,\dots, K_2. \end{aligned}\] Finally, since the ABA equations in the classical regime, \(h\rightarrow \infty\), \(L\sim K_\alpha\rightarrow \infty\) with fixed \(\mathcal{L}\equiv L/h\) reproduce the classical limit ([\[eq:classical1\]](#eq:classical1){reference-type="ref" reference="eq:classical1"})-([\[eq:classical4\]](#eq:classical4){reference-type="ref" reference="eq:classical4"}) via condensation of roots into cuts in the standard way , it follows that we also reproduce the classical limit from the QSC, similarly to. Thus we see that our QSC successfully reproduces all the data from section [2](#sec:data){reference-type="ref" reference="sec:data"}.
# Discussion and outlook
The QSCs for \(AdS_5\) and \(AdS_4\) have a lot in common with one another--both are based on QQ-relations dictated by the global symmetries and have similar additional analyticity constraints. We use these general features to propose a QSC for string theory on \(AdS_3 \times S^3 \times T^4\) with RR charge and its \(\text{CFT}_2\) dual. However, in contrast to the higher-dimensional QSCs, the assumption of square-root singularity near the branch points needs to be dropped. While we reproduced successfully the ABA equations for massive modes, we should still emphasise that, unlike in the previous cases , we do not have the luxury of TBA equations which can be used as a starting point to derive the QSC equations. Instead, we use a bottom-up approach where we guess the QSC based on the symmetries and analogy with previous cases, and then verify it in some limits. On the important point of the order of the cuts, we notice that if we assume them to be of the usual square-root type, unlike in previous cases, we get a further nontrivial algebraic constraint on the Q-functions in addition to the QQ-relations, resulting in a too restrictive set of equations. So to some extent the absence of square-root behaviour is dictated by symmetries. To be fully confident in the self-consistency and completeness of equations proposed here, we need to perform further tests beyond the matching with the ABA presented here. For example, constructing the perturbative weak coupling solution at several loop orders would be useful, which can be done with the methods of. The QSC should also reproduce the protected spectrum of the theory , accounting for all-order wrapping corrections not considered in the ABA analysis . Further, it would be interesting to consider near-BPS limits where one can expect a non-trivial analytic solution at finite coupling . Finally, one should try to solve the system numerically with high precision like in. Another potential way to test our equations would be to re-derive the TBA equations for the massless modes. An important question to address is whether the massless modes of the theory are already contained in our QSC proposal or whether the construction needs to be generalised in some way. For example, one might wonder whether it is possible to take a tensor product of our Q-system with an additional Q-system, perhaps based on \(\alg{su}(2)_\circ\) under which the massless bosons are known to transform . Unfortunately, such a direct product is not compatible with the fact that massless fermions transform non-trivially under the \(\alg{psu}(1,1|2)^2\) symmetry. On the other hand, the structure of the Q-system is rather rigid so it is harder to see how to augment it by an additional Q-system in a more non-trivial way. Alternatively, it may be that incorporating the massless modes requires relaxing slightly some of the analyticity and pole-structure properties we require of the QSC. The starting point for this approach would be an attempt to derive the ABA equations including massless modes in the large-volume limit, generalising the construction we presented above. This would require new arguments, since the notion of the asymptotic regime is delicate in the presence of massless excitations, where the standard exponential large-\(L\) suppression is no longer there and wrapping corrections often are of the same order as the ABA contributions. However, we point out that the QSC structure is typically very rigid and does not allow for much more freedom. In particular, given the underlying \(\alg{psu}(1,1|2)\) algebraic structure of our system, we believe that the only place where the QQ relations could be changed would be a modification of the condition \(Q_{12|12}=Q_{\dot 1\dot 2|\dot 1\dot 2}=1\). We might also have to modify the gluing condition, although this seems less natural. However, we cannot at this stage exclude the possibility that neither of these options will be necessary and the QSC presented here is already complete. To provide evidence for this claim, it would be very interesting to solve the QSC equations at finite coupling, as this would allow to perform tests in various limits, for example comparing with massless solutions at weak coupling or in the semi-classical regime . If these additional tests can be satisfactorily performed, one can hope that \(AdS_3\) would become an ideal background for application of SoV program for correlators . Further, combining the \(AdS_3\) QSC spectral methods with Conformal Bootstrap techniques could provide a simpler testing ground for these ideas compared to the \({\cal N}=4\) SYM case. Following these tests of our conjecture, it would be interesting to extend the \(AdS_3\) QSC construction to \(AdS_3 \times S^3 \times T^4\) backgrounds supported by combinations of RR and NSNS charges. The ABA for these theories is also known and solutions to the crossing equations have recently been found , which should provide a further testing ground for the QSC analysis. String theory on \(AdS_3 \times S^3 \times S^3\times S^1\) is also expected to be integrable . Finding the QSC for this model would be particularly interesting since the global symmetry algebra is \(d(2,1;\alpha)^2\), for which the Q-system should exhibit novel features. It would be interesting to see whether similar techniques to the ones we have employed here can be extended to the \(AdS_2/CFT_1\) integrable system , which also features the presence of massless modes and has an algebraic structure of a similar complexity. The issue of long vs short representations, which is relevant in that case, is likely to represent an additional novelty and a reason for adapting the method even further. | {'timestamp': '2022-08-09T02:23:48', 'yymm': '2109', 'arxiv_id': '2109.05500', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05500'} |
null | null |
null | null |
null | null |
# Introduction
Sickle cell disease (SCD) is a disorder caused by mutations in position 6 of \(\beta\)-hemoglobin gene (hemoglobin S) with extreme variability at phenotypic level. In some patients the disease manifestation is so mild that they remain asymptomatic most of the time while others die before the age of five from several of the severe complications associated to the SCD. Individuals who have the hemoglobin S variant are naturally protected against malaria, which has a profound influence on the spread of sickle cell disease globally affecting the tropical (African and Asian) countries the most. Many of these countries are not able to support diagnosis and appropriate healthcare for this group of patients leading to a drop in the life expectancy from 45-55 years in high income countries to 90% death rate before the age of 5 in low income countries. Severity monitoring and prediction of the SCD is therefore an important task along with development of new effective and inexpensive therapeutic strategies. Changes in severity allow monitoring of the treatment efficiently and for prediction and prevention of life-threatening complications in short future. To date, there is no practical test based on red blood cells (RBCs) density separation analysis available for prediction of the severity of the disease for a patient. An important parameter for disease severity assessment is the percentage of hypo-and hyperchromic cells. RBCs with a hemoglobin concentration above 410 g/l are called hyperchromic and characterized by low cellular deformability and increased probability of aggregation of the hemoglobin S which directly associates with advanced severity and poor prognosis for the SCD patients. In contrary, Hypochromic RBCs with low hemoglobin content are associated with a lower probability of hemoglobin S aggregation and sickling and thus with mild disease manifestation. Measurement of these parameters using blood smears is laborious and time-consuming, when done manually by skilled personnel, or rather requires expensive medical laboratory equipment, when automated. The spleen plays an important role in clearing the blood from old, broken, dehydrated or hyperchromic red blood cells (RBCs). A normal and functioning spleen reduces the intravascular hemolysis of damaged cells (where cells rupture in the blood vessels) and prevents vaso-occlusive crisis (where terminally dense sickle cells block circulation of blood vessel leading to painful crisis) and vascular damage in SCD patients. However, fibrosis and progressive atrophy of the spleen resulting finally in necrosis of the organ, known as autosplenectomy, which is often observed in SCD patients with severe disease phenotype. It is a known problem in children with SCD due to repeated splenic vaso-occlusive events in the organ. Measuring spleen size with ultrasound is a common way to evaluate the organ's condition in SCD. We here propose a computational approach that circumvents expensive lab tests and relies solely on the measurement of spleen size and a Percoll image. Percoll images are used to assess the density of the cells and particles. After centrifugation, several bands with different thicknesses are formed by RBCs of similar density (see Fig. [\[overview\]](#overview){reference-type="ref" reference="overview"}) holding important information about a SCD patient's condition. Back in 1984, Fabry et al. observed a decrease in the dense fraction of Percoll images in SCD patients suffering from painful crisis in 11 patients over 14 painful crisis image. This information can also be computationally analyzed: Sadafi et al. introduced a hybrid approach based on CNNs and features extracted from fast Fourier transform to classify a variety of hereditary hemolytic anemias using Percoll image data. To predict the severity of the SCD patients, we are proposing an approach based on graph convolutional networks (GCN) to form a population graph on our data. The similarity of the GCN edges is calculated using lab (Percentage of hypo-and hyperchromic RBCs) and clinical data (spleen size). The spleen size is measured using ultrasound. We propose a CNN based approach to have an easy to access and affordable anywhere in the world way to estimate required lab data from the Percoll image.
# Methodology
Our proposed method, SCD-severity-GCN aims at predicting SCD severity from cheap and easy accessible patient data and consists of the following steps: (i) The abundance of hypo-and hyperchromic cells in the blood sample are predicted based on a Percoll image; (ii) Relevant features are extracted from the Percoll image using a CNN and fast Fourier transform (FFT). (iii) A similarity metric between Percolls based on a patient's spleen size and the predicted abundance of hypo-and hyperchromic cells is calculated to form a population graph. Using GCNs the SCD severity is predicted (Fig [\[overview\]](#overview){reference-type="ref" reference="overview"})
## Model
Our goal is to have a model \(f\) that takes a Percoll image \(P_i\) and the spleen size \(J_i\) of a patient sample \(i\) to return a severity grade \(S_i\): \[S_i = f(P_i, J_i; \theta)\] where \(\theta\) are the model parameters that are learned by training on the dataset.
## Feature extraction
For primary feature extraction the approach proposed in is employed. There, the extraction of Fourier features from the images has been demonstrated to enhance disease classification performance on Percoll images. Accordingly, we extract features with an AlexNet architecture and combine them with features from FFT (see Fig. 1). We obtain pretrained weights of the model \(f_{\mathrm{cnn-fft}}\) and use the activations preceding the final classification layer as features for our GCN approach. To reduce feature dimensions, we used recursive feature elimination and a Ridge classifier as suggested by Parisot et al.. \[x_i = \mathrm{RFE}(f_{\mathrm{cnn-fft}}) (P_i)\] where \(x_i\) is the feature vector extracted for the Percoll image \(P_i\). Also in our approach this step improved the convergence of the training significantly.
## Graph convolution network
One of the most intuitive ways of representing populations and their similarities is through graphs. In our approach, every Percoll image \(P_i\) is represented by a vertex \(v \in \mathcal{V}\) and the similarity between the Percoll images is modelled by weighted edges \(\mathcal{E}\) calculated from the expensive laboratory data (the percentages of hypo-and hyperchromic) which are predicted and cheap clinical data (i.e. spleen size of the patient) (see Fig 2). A population graph \(\mathcal{G} = \{\mathcal{V},\mathcal{E}\}\) is defined accordingly.
## Hemoglobin density estimation
To allow for an application of the method without expensive laboratory testing, the percentages of hypo-and hyperchromic cells \(\hat{H}\) in the blood are estimated by a regression. A CNN \(f_{chrome}\) is proposed for this task. The groundtruth values \(H\) are provided for every Percoll image and are used to train the network: \[\mathcal{L}_{chrome} (\gamma) = \frac{1}{N} \sum_{i = 1}^{N} (H_i-\hat{H_i})^2\] where \(\hat{H_i} = f_{chrome}(P_i; \gamma)\) and \(\gamma\) is the network parameters.
## Similarity metric
Under the assumption that patients with similar features experience comparable severity of the disease, the similarity between two samples \(v\) and \(w\) is calculated via \[\mathcal{E}(v_i,v_j) = e^{-(||\hat{H}_{v_i}-\hat{H}_{v_j}|| + \lambda[J_{v_i} == J_{v_j}])}\] where \(\hat{H}\) is the vector of estimated percentages of hypo-and hyperchromic cells and \(J\) is the spleen size, as above. Iverson brackets yield 1 in case of equality and 0 otherwise. Note that spleen sizes are given as discrete numbers in centimeters (see Fig. 2), obtained in the clinic with a conventional ultrasound device. The coefficient \(\lambda\) is set to weight the importance of spleen and lab measurements.
# Experiments
## Dataset
Our dataset consists of the 216 samples with Percoll images and laboratory data (% hypo, %hyper) and clinical data (spleen size) obtained from 17 patients diagnosed with SCD, who participated in a clinical trial (NCT03247218) conducted in Emek Medical Center in Afula[^1]. The study has been conducted in accordance with local ethics committee guidelines and the Declaration of Helsinki. Blood samples were acquired during pre-planned monthly visits according to the trial protocol. For every visit the patient's health was evaluated using blood analysis, including RBC characteristics and measurement of hemolytic and inflammatory markers, urine analysis and blood pressure measurements. Severity of a patient's condition at each measurement point was estimated using the scoring approach proposed by Sebastiani et al. with minor modifications on disease severity score calculation. Figure [\[fig.dataset\]](#fig.dataset){reference-type="ref" reference="fig.dataset"} shows distribution of severity scores and example samples from the dataset.
## Implementation details
**Hemoglobin density estimation**: A CNN with seven convolutional layers with ReLU activation function and max-pooling is used. After global average pooling and two fully connected layers the output is regulated with a final ReLU. Two dropout layers with a drop rate of \(0.5\) are used for regularization.\
**Feature extraction**: The output size from CNN-FFT is 1024, which is reduced with recursive feature elimination (RFE) to 50 features. These features are used as the final feature vector for each Percoll image.\
**Graph convolutional network**: A population graph is created based on the defined feature vectors and similarities. We use two hidden layers in the graph and 50 filters in the hidden layers. The dropout rate is set to \(0.2\). For similarity calculation \(\lambda\) is set to \(10\).\
**Training**: Both training procedures are carried out on a 10-fold cross validation dataset. The model \(f_{chrome}\) estimating hemoglobin density is trained with AMSGrad variation of Adam optimizer for \(100\) epochs and a learning rate of \(0.0005\). The graph convolutional network is trained for 300 epochs using Adam optimizer and a learning rate of \(0.01\). We use the Tensorflow framework for implementation and training.\
**Evaluation metrics**: We are reporting root mean square error (RMSE) for the regression task of hemoglobin density estimation. Accuracy, weighted F1-score and area under ROC are reported for the severity grading as well as the area under precision recall curve for every class. Scikit-learn implementation is used for calculation all of the metrics.\
**Baseline**: A linear SVM trained on the feature vectors is used as a baseline for our grading approach.
## Results
The dataset is divided into 10 stratified folds for patient-wise cross validation. All of the models are independently ran on each combination of these folds. Mean and standard deviation is reported for all of the 10 experiments. First, the values predicted by the Hb density estimation model \(f_{chrome}\) based on Percoll images are compared against the actual lab tests. The root mean square error (RMSE) of the percentage for hypochromic cells is \(6.5 \pm 4.0\) and for hyperchromic cells \(0.90 \pm 0.12\). Considering the ranges of the hypo and hyper values, which are \([0.6, 37.5]\) and \([0.2, 10.9]\), respectively, we consider the estimation sufficiently good. Next, we compare our SCD-severity-GCN approach with the following methods: (i) A linear SVM trained on the \(x_i\) features vectors extracted from the Percoll image (SVM), (ii) a linear SVM trained on \(x_i\) feature vectors and the cheap clinical ultrasound and newly proposed and time consuming groundtruth lab information (SVM-Lab), (iii) a GCN based on randomized laboratory information (GCN-Rand), and (iv) a GCN using not the estimated, but the actual laboratory information (GCN-Lab) as the upper limit. Table [\[tab.results\]](#tab.results){reference-type="ref" reference="tab.results"} shows that our SCD-severity-GCN approach using estimated Hb densities is close to the GCN that required hard to obtain lab data (GCN-Lab) in terms of accuracy, weighted F1-score and area under ROC. Since the dataset is unbalanced, we are reporting the area under precision recall curve in Figure [\[fig_auprc\]](#fig_auprc){reference-type="ref" reference="fig_auprc"} for every class and different approaches.
## Ablation study
GCNs are sensitive to the formulation of the graph adjacency matrix based on the pairwise similarity that is defined between the nodes. Choosing parameters that are biologically significant and easy to obtain is crucial. To evaluate the importance of the different clinical (spleen size) and laboratory (% of hypo-and hyperchromic cells) information used for the formation of our GCN, we designed an ablation study and compare GCNs trained with different combinations of these parameters. As Table 2 shows, the combination based on spleen size and percentages of hypo-and hyperchromic RBCs yields the best result.
## Discussion
Severity prediction of SCD is a challenging task normally preformed with several clinical and laboratory tests. Here we propose a novel potential severity prediction approach based on RBC density separation (as provided by Percoll gradients) analysis that may amend the currently existing ones. Information obtained solely from Percoll images is not be sufficient for an acceptable classification (see Table [\[tab.results\]](#tab.results){reference-type="ref" reference="tab.results"}), even though those features sufficed for successful diagnosis of different anemias. By combining Percoll derived features with complementary clinical and laboratory data and training a GCN with this information, we can achieve an accuracy that is surprisingly high for this challenging clinical task. This is illustrated by the UMAP embedding of feature vectors (Fig [\[fig.umap\]](#fig.umap){reference-type="ref" reference="fig.umap"} a), and GCN outputs with estimated (Fig [\[fig.umap\]](#fig.umap){reference-type="ref" reference="fig.umap"} b) and groundtruth lab information (Fig [\[fig.umap\]](#fig.umap){reference-type="ref" reference="fig.umap"} c). Samples from different severity classes are nicely disentangled in the UMAP thanks to the GCN approach we utilized. Although clustering using the groundtruth lab information (GCN-Lab) is a lot better, a smooth transition from low to high severity is already evident in the approach that uses estimated Hb density only (SCD-severity-GCN).
# Conclusion
Sickle cell disease severity prediction is an important task that allows to prevent life-threatening complications, reduce morbidity and mortality and refine the choice of optimal therapeutic strategies. Offering affordable and versatile solutions for improving life quality of the SCD patients is a necessity specially in low resource areas of the planet. Here, we proposed the first computational method requiring only the Percoll gradient image and spleen size obtained from a conventional ultrasound. Analysis of Percoll gradient images with CNNs nicely predicted percentages of hypo-and hyperchromic cells and the proposed GCN predicted SCD severity score with a surprisingly high accuracy. Our approach uses a unique combination of methods, with a GCN at its heart. Results look very promising and provide a solid ground for future work. Next we will analyze more patients, especially more severe ones as well as pediatric datasets. Our SCD-severity-GCN based on Percoll images requires much smaller volumes of blood compared to common hematological tests (1 ml or less instead of 7-10 ml), which is particularly relevant for kids and patients suffering from severe anemia. | {'timestamp': '2021-09-14T02:17:08', 'yymm': '2109', 'arxiv_id': '2109.05372', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05372'} |
# Introduction
Throughout, we will let \(\{ x \}, \norm{x}\) denote the fractional part of \(x\) and the distance from \(x\) to the nearest integer respectively. A sequence of reals \(\left( x_n \right)_{n \geq 1}\) is *equidistributed modulo 1* if for all \(0 < a < b < 1\) we have \[\lim_{N \to \infty} \frac{1}{N} \sum_{n = 1}^{N} \mathbbm{1} \left( \{ x_n \} \in [a, b] \right) = b-a\] that is the number of \(x_n\)'s which lie in a certain interval is asymptotically proportional to the length of that interval. A sequence of reals \(\left( x_n \right)_{n \geq 1}\) is said to have *Poissonian pair correlation* if for all \(s > 0\) we have \[\lim_{N \to \infty} \frac{1}{N} \sum_{1 \leq m \neq n \leq N} \mathbbm{1} \left( \norm{x_m-x_n} < \frac{s}{N} \right) = 2 s\] which essentially means that the local spacing at scale \(\frac{1}{N}\) between pairs of elements is the same as in the random case. Poissonian pair correlation implies equidistribution (see, ), and so it can be thought as a finer, more local statistic. Given a sequence of reals \(x_n\), a natural number \(N > 0\) and a real \(\gamma > 0\), define the additive energy at scale \(\gamma\) to be \[E_{N, \gamma} = \# \left\{ 1 \leq a, b, c, d \leq N \ \big\lvert \ \abs{x_a + x_b-x_c-x_d} < \gamma \right\}\] and when the scale is not written we let \(E_{N} = E_{N, 1}\). We always have \(E_{N, \gamma} \geq N^2\) by the contribution of the diagonal \(a = c, b = d\). If for example we are given a spacing condition like \(x_{n + 1}-x_n \geq \delta\) then \(E_{N, \gamma} \leq \frac{2 \gamma N^3}{\delta} + N^3\) as \(a, b, c\) define \(d\) up to \(\frac{2 \gamma}{\delta} + 1\) options. We say that an increasing sequence of real numbers \(\left( x_n \right)_{n \geq 1}\) is *well spaced* if \(x_{n + 1}-x_n \geq 1\). Given a property \(T\) of sequences, we say that a sequence \(\left( x_n \right)_{n \geq 1}\) has the property *metric \(T\)* if for almost all \(\alpha \in \mathbb{R}\) in the sense of Lebesgue measure, the sequence \(\left( \alpha x_n \right)_{n \geq 1}\) has the property \(T\). Aistleitner, El-Baz and Munsch in raised the problem of whether there exists an increasing sequence of reals \(\left( x_n \right)_{n \geq 1}\) with additive energy of high order, that is \(E_{N} \gg N^3\), and metric Poissonian pair correlation. Recently, Lutsko, Sourmelidis and Technau have shown that for all \(0 < \theta < \frac{1}{3}\) and for all \(\alpha \neq 0\) the sequence \(\alpha n^{\theta}\) has Poissonian pair correlation, and Rudnick and Technau have shown that for all \(0 < \theta < 1\), the sequence \(\left( n^{\theta} \right)_{n \geq 1}\) has metric Poissonian pair correlation, answering this question in the affirmative. However, as opposed to most of the sequences discussed in, these sequences are not well spaced, that is \(x_{n + 1}-x_n \to 0\). We will show that there exist well spaced sequences, with additive energy \(E_{N} \gg N^3\) and metric Poissonian pair correlation. In it is conjectured that an increasing sequence of reals with additive energy \(E_{N, \gamma} \gg \gamma N^4\) where \(\frac{1}{N} \leq \gamma = \gamma (N) \leq 1\) is some function of \(N\) is sufficient to ensure lack of Poissonian pair correlation. In the case of well spaced sequences, then for all \(\gamma \leq 1\) we have \(E_{N, \gamma} \ll N^3\), and so this criterion can only possibly hold when \(\gamma \ll \frac{1}{N}\). The analogous question one can ask is given a well spaced sequence such that \(E_{N, \gamma} \gg N^3\), how small must \(\gamma\) be to ensure that \(x_n\) does not have metric Poissonian pair correlation. More specifically, we ask\
\
**Open Problem 1:** Show that a well spaced sequence with additive energy \(E_{N, \gamma} \gg N^3\) does not have metric Poissonian pair correlation, where \(\gamma = \frac{1}{N}\).\
\
In the case where \(x_n\) is an integer valued sequence, notice that \(E_{N, \gamma} = E_N\) for all \(\gamma \leq 1\). In an appendix to a paper by Aistleitner, Larcher, and Lewko, Bourgain showed that if \(x_n\) is an increasing sequence of integers with \(E_N \gg N^3\) for infinitely many \(N\), then \(x_n\) does not have metric Poissonian pair correlation. This argument was later sharpened by Lachmann and Technau, and essentially optimized by Larcher and Stockinger, who showed that in this case, there is no \(\alpha\) such that the sequence \(\alpha x_n\) has Poissonian pair correlation. We can adapt the argument of Bourgain to the real valued case, by defining \(y_n = \frac{1}{C N} \floor{C N x_n}\) for some large constant \(C\). Then, the large additive energy of \(x_n\) with respect to a small scale transfers to many equalities of the form \(y_a + y_b = y_c + y_d\), which then allows us to use the Balog-Szemeredi-Gowers theorem. This follows through to show that for all \(c_1 > 0\) there is a constant \(c_2 > 0\) such that if \(E_{N, \frac{c_2}{N}} \geq c_1 N^3\) then the sequence \(x_n\) does not have metric Poissonian pair correlation, but we have not managed to quite push it all the way to \(E_{N, \frac{1}{N}} \gg N^3\). In the other direction, we show the following:
Our construction is probabilistic. Recall that the discrepancy of a sequence \(x_n\) of numbers modulo 1 is defined as \[D_N = \sup_{0 \leq a \leq b \leq 1} \abs{\frac{\# \left\{ n: x_n \in \left[ a, b \right] \right\}}{N}-\left( b-a \right)}\] Discrepancy is a quantitative version of equidistribution: the more evenly distributed a sequence is modulo 1, the smaller the discrepancy. Define \(M_N = \max_{n \leq N} n D_n\). We will prove the following statement:
Let us just mention that the regularity conditions on \(g\) are just a technicality, and in all reasonable cases they are satisfied. A result of Schmidt states that there is some constant \(c > 0\), such that infinitely often we have \(D_N \geq \frac{c \log N}{N}\), and therefore it is reasonable to assume that \(g(N) \geq \frac{\log N}{N}\) which is stronger than the first regularity condition. As for the second condition: intuitively, \(\frac{M_N}{N}\) should be about \(D_N\), and then \(\frac{D_N}{g(N)} \to 0\), and so all we require is that when we multiply \(N\) by a number which is \(1 + o(1)\), the value of the function \(g\) should change by \(1 + o(1)\), which happens for most reasonable options for \(g\). This regularity condition can be slightly weakened further, but for the reasons stated we have not attempted to optimize it. This theorem suggests that for a \"typical\" sequence, Poissonian pair correlation (or other local statistics) are in some sense a \"smoothed out\" version of small discrepancy, and it is evident from the proof as well: the small discrepancy ensures that we are evenly distributed on a relatively small scale, and the random shifts by \(z_n\) smooth out the sequence to give us pseudorandom behaviour at a scale of \(\frac{1}{N}\), that is Poissonian pair correlation. The connection between these two statistics has been studied for example by Steinerberger who has shown that a sequence that has a certain uniform version of Poissonian pair correlation has small discrepancy. Interestingly, Larcher and Stockinger have established that many classical low discrepancy sequences do not have Poissonian pair correlation. Let us show the deduction of Theorem [\[real-result\]](#real-result){reference-type="ref" reference="real-result"} from Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"}: Let \(g(N) = \frac{\log N \left( \log \log N \right)^{1 + c}}{N}\). We will show that almost surely, a sequence of the form \(x_n = 2 n + z_n\), where \(z_n \sim \mathrm{Unif} \left[-g(N), g(N) \right]\) are independent random variables, has metric Poissonian pair correlation. Clearly, this sequence satisfies \(x_{n + 1}-x_n \geq 1\). This sequence has additive energy \(E_{N, \gamma} \gg N^3\), where \(\gamma = 10 g(N)\), because there are \(\gg N^3\) quadruples \(\frac{N}{2} \leq a, b, c, d \leq N\) such that \(a + b = c + d\), and for each such quadruple we have \(\abs{x_a + x_b-x_c-x_d} \leq 10 g(N)\). Now, by Fubini's theorem, in order to show that almost surely, for almost all \(\alpha\) the sequence \(\alpha x_n\) has Poissonian pair correlation, it is sufficient to show that for almost all \(\alpha\), almost surely the sequence \(\alpha x_n\) has Poissonian pair correlation. By a theorem of Beck (though this specific case was proven first by Khintchine ), for almost all \(\alpha\), the sequence \(\alpha n\) has discrepancy \[D_N = \mathcal{O} \left( \frac{\log N \left( \log \log N \right)^{1 + c/2}}{N} \right)\] and a direct application of Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"} gives us the desired result. In Beck showed that for almost all \(\alpha\), for infinitely many \(N\) we have that the discrepancy of the sequence \(\alpha n\) is at least \(D_N \geq \frac{\log N \log \log N}{N}\), and so Theorem [\[real-result\]](#real-result){reference-type="ref" reference="real-result"} is the best we can get from Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"} with a sequence of the form \(x_n = 2 n + z_n\). This raises a natural question, which to the best of our knowledge seems not to have been discussed in the literature:\
\
**Open Problem 2:** Is there an increasing sequence of integers \(x_n\) such that for almost all \(\alpha\), the discrepancy of \(\alpha x_n\) is \(D_N \ll \frac{\log N \log \log N}{N}\)?\
\
It seems reasonable to conjecture that the answer to this question is negative: for example, Aistleitner and Larcher have shown that an increasing sequence of integers \(x_n\) such that for almost all \(\alpha, \ \alpha x_n\) has small discrepancy, must have large additive energy, which is a significant structural condition on the sequence. As we have stated above, Schmidt has shown that for any sequence modulo 1, for infinitely many \(N\), the discrepancy is at least \(D_N \geq \frac{\log N}{N}\), and so the best improvement to Theorem [\[real-result\]](#real-result){reference-type="ref" reference="real-result"} that our method involving Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"} could theoretically yield is \(\gamma = f(N) \frac{\log N}{N}\), where \(f\) is a slowly growing function which tends to infinity with \(N\). Therefore, we ask\
\
**Open Problem 3:** Is there a well spaced sequence \(x_n\) with metric Poissonian pair correlation, with additive energy \(E_{N, \gamma} \gg N^3\), where \(\gamma \ll \frac{\log N}{N}\)?\
\
The proof of Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"} breaks down in this case if we try to apply the same method, which hints maybe that such sequences (if they exist) are atypical in some sense. We show a certain converse theorem to [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"}, which is not tight, but at least shows that our method in Theorem [\[real-result\]](#real-result){reference-type="ref" reference="real-result"} cannot get metric Poissonian pair correlation and additive energy at a scale of \(\frac{\sqrt{\log N}}{N}\).
For an integer \(k \geq 2\), \(k\)-level correlation is a statistic which intuitively measures the local spacing at scale \(\frac{1}{N}\) between \(k\)-tuples of elements. More formally, let \(\mathcal{X}_k = \mathcal{X}_k (N)\) denote the set of \(k\)-tuples \(\left( a_1, \dots, a_k \right)\) of distinct integers \(1 \leq a_i \leq N\). Given \(s = \left( s_1, \dots, s_{2 k-2} \right) \in \mathbb{R}^{2 k-2}\) where \(s_{2 i-1} < s_{2 i}\) for all \(1 \leq i < k\), let \(\mathbbm{1}_{s, N} \left( y_1, \dots, y_{k-1} \right)\) denote the indicator function of the event \(\{ y_1 \} \in \left[ \frac{s_1}{N}, \frac{s_2}{N} \right], \dots, \{ y_{k-1} \} \in \left[ \frac{s_{2 k-3}}{N}, \frac{s_{2 k-2}}{N} \right]\) where the intervals are taken modulo 1, that is for our purposes \(\frac{3}{4} \in \left[-\frac{1}{3}, \frac{1}{3} \right]\). A sequence of reals \(\left( x_n \right)_{n \geq 1}\) is said to have *Poissonian \(k\)-level correlation* if for all \(s_1, s_2, \dots, s_{k-1} > 0\) we have \[\lim_{N \to \infty} \frac{1}{N} \sum_{a \in \mathcal{X}_k} \mathbbm{1}_{s, N} \left( x_{a_1}-x_{a_2}, x_{a_1}-x_{a_3}, \dots, x_{a_1}-x_{a_k} \right) = \left( s_2-s_1 \right) \left( s_4-s_3 \right) \cdots \left( s_{2 k-2}-s_{2 k-3} \right)\] Analogously to pair correlation, Poissonian \(k\)-level correlation means that the local spacing at scale \(\frac{1}{N}\) between \(k\)-tuples of elements is the same as in the random case. Usually in the literature Poissonian \(k\)-level correlation is defined by the property that for any smooth, compactly supported \(f \in C_{c}^{\infty} \left( \mathbb{R}^{k-1} \right)\) we have \[\lim_{N \to \infty} \frac{1}{N} \sum_{a \in \mathcal{X}_k} \sum_{m \in \mathbb{Z}^{k-1}} f \left( N \left( m + \left( x_{a_1}-x_{a_2}, x_{a_2}-x_{a_3}, \dots, x_{a_{k-1}}-x_{a_k} \right) \right) \right) = \int_{\mathbb{R}^{k-1}} f(x) \mathrm{d} x\] This is equivalent to our definition, by first defining \(g \left( y_1, \dots, y_{k-1} \right) = f \left( y_1, y_1 + y_2, \dots, y_1 + \cdots + y_{k-1} \right)\) which is also smooth and compactly supported, and in one direction approximating \(g\) from above and below by a linear combination of indicators of boxes, and in the other direction approximating an indicator function of a box from above and below by a smooth test function. Theorem [\[real-result\]](#real-result){reference-type="ref" reference="real-result"} generalizes in a straightforward manner in the following way:
This follows in the same manner from a generalization of Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"}:
The proof is very similar to the proof of Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"}, and so we will just sketch what the differences are and how to overcome them. The main reason for studying higher level correlations is that the \(k\)-level correlations for all \(k \geq 2\) essentially determine the local statistics entirely. More precisely, let \(x_{(1)}^{N} \leq x_{(2)}^{N} \leq \cdots \leq x_{(N)}^{N}\) denote the first \(N\) elements of the sequence \(x_n\), ordered by size, and define \[h \left( x, \left( x_n \right)_{n \geq 1}, N \right) = \frac{1}{N} \# \left\{ 1 \leq n < N \: \ N \left( x_{(n)}^{N}-x_{(n + 1)}^{N} \right) \leq x \right\}\] The *level spacing distribution* (also called the *gap distribution*, or the *nearest neighbour spacing distribution*) is the limiting distribution \(P(s)\) (if it exists), which satisfies \[\lim_{N \to \infty} h \left( x, \left( x_n \right)_{n \geq 1}, N \right) = \intop_{0}^{x} P(s) \mathrm{d} s\] The level spacing distribution contains the majority of local information at scale \(\frac{1}{N}\) of the sequence \(x_n\), and so it is naturally an object of interest. \(N\) random points on the unit circle approximate a Poisson point process, and so in the random case the level spacing distribution is asymptotically a Poisson distribution, that is \(P(s) = e^{-s}\). Accordingly, a sequence is said to have *Poissonian level spacing distribution* if its level spacing distribution is \(P(s) = e^{-s}\). It is well known that Poissonian \(k\)-level correlation for all \(k \geq 2\) implies Poissonian level spacing distribution. Thus, [\[k-level\]](#k-level){reference-type="ref" reference="k-level"} and [\[disc-k\]](#disc-k){reference-type="ref" reference="disc-k"} immediately imply
**Acknowledgements**: The authors thank Zeev Rudnick for helpful comments, in particular pointing out Corollary [\[arithmetic_cor\]](#arithmetic_cor){reference-type="ref" reference="arithmetic_cor"}. The second author received funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (Grant agreement No. 786758).
# Proof of Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"}
Define the random variable \[X_{s, N} = \frac{1}{N} \sum_{1 \leq n \neq m \leq N} \mathbbm{1}_{s, N} \left( x_n + z_n-x_m-z_m \right)\] By definition, we have Poissonian pair correlation if and only if almost surely \[\lim_{N \to \infty} X_{s, N} = 2 s\] It is enough to show this for each \(s\) individually, as if this is true for all \(s \in \mathbb{Q}\) then it is true for all \(s\), and a countable intersection of sets of full measure is a set of full measure. For simplicity, from here on we will denote \(X_N = X_{s, N}\). We will prove this via a second moment method. First, we compute the expectation of \(X_N\). \[\mathbb{E} \left[ X_N \right] = \frac{1}{N} \sum_{1 \leq m \neq n \leq N} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_n + z_n-x_m-z_m \right) \right]\] For \(x \in \mathbb{R}\), denote \[\rho_N (x) = \frac{1}{N} \sum_{n = 1}^{N} \frac{1}{2 g(n)} \mathbbm{1} \left( x \in \left[ x_n-g(n), x_n + g(n) \right] \right) = \frac{1}{N} \sum_{n = 1}^{N} \frac{1}{2 g(n)} \mathbbm{1} \left( x_n \in \left[ x-g(n), x + g(n) \right] \right)\] One should think of \(\rho_N (x)\) as the density function of the probability measure which corresponds to balls around \(x_n\) of radius \(g(n)\), each with a weight of \(1/N\), that is for every function \(f\) we have \[\frac{1}{N} \sum_{n = 1}^{N} \mathbb{E} \left[ f \left( x_n + z_n \right) \right] = \intop_{0}^{1} f \left( x \right) \rho_N (x) \mathrm{d} x\] Now, we define \[h_{s, N} (x) = \sum_{n = 1}^{N} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_n + z_n-x \right) \right]\] By using linearity of expectation, \[\mathbb{E} \left[ X_N \right]-\frac{1}{N} \sum_{m = 1}^{N} \mathbb{E} \left[ h_{s, N} \left( x_m + z_m \right) \right] = \frac{1}{N} \sum_{m = 1}^{N} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( z_m-z_{m}' \right) \right]\] where \(z_{1}', z_{2}', \dots\) is an independent copy of \(z_1, z_2, \dots\). Clearly, \[\mathbb{E} \left[ \mathbbm{1}_{s, N} \left( z_m-z_{m}' \right) \right] \leq \frac{s}{N g(m)} \leq \frac{s}{N g(N)}\] and so \[\frac{1}{N} \sum_{m = 1}^{N} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( z_m-z_{m}' \right) \right] \leq \frac{s}{N g(N)} = o_N \left( 1 \right)\] Therefore, \[\mathbb{E} \left[ X_N \right] = \intop_{0}^{1} h_{s, N} (x) \rho_N (x) \mathrm{d} x + o_N \left( 1 \right)\] What we need to show is that \(\intop_{0}^{1} h_{s, N} (x) \rho_N (x) \mathrm{d} x\) converges to \(2 s\) as \(N \to \infty\). We will do this by showing that uniformly in \(x\), we have \(\rho_N (x) \to 1, \ h_{s, N} (x) \to 2 s\). Let us start with \(\rho_N (x)\). This will follow from the fact that the discrepancy of \(x_n\) is much smaller than \(g(n)\): essentially, \(\rho_N(x)\) counts the number of points \(x_n\) in a certain interval which is of a length much larger than the discrepancy, and so we have the correct asymptotic for the number of \(x_n\)'s in that interval. There is a slight complication, due to the fact that the interval is not fixed. To remedy this, we will divide our sum into intervals of the form \(L \leq n \leq L + k(L)\) for some \(L, k(L)\), where we need to ensure simultaneously that \(g\) is approximately the same at both endpoints of the range of summation (which allows us to treat the interval \(\left[ x-g(n), x + g(n) \right]\) as fixed), and that the range of summation is long enough so that we can use our discrepancy bounds. Specifically, we will take \(k(L)\) small enough such that \(\frac{g \left( L + k(L) \right)}{g(L)} \to_{L \to \infty} 1\). Then, \[\sum_{L < n \leq L + k(L)} \frac{1}{2 g(n)} \mathbbm{1} \left( x_n \in \left[ x-g(n), x + g(n) \right] \right) \sim \frac{1}{2 g(L)} \sum_{L < n \leq L + k(L)} \mathbbm{1} \left( x_n \in \left[ x-g(n), x + g(n) \right] \right)\] Because \(g\) is monotonically decreasing, \[\mathbbm{1} \left( x_n \in \left[ x-g \left( L + k(L) \right), x + g \left( L + k(L) \right) \right] \right) \leq \mathbbm{1} \left( x_n \in \left[ x-g(n), x + g(n) \right] \right) \leq \mathbbm{1} \left( x_n \in \left[ x-g(L), x + g(L) \right] \right)\] Letting \(c\) be either \(g(L)\) or \(g \left( L + k(L) \right)\), we will show that \[\frac{1}{2 g(L)} \sum_{L < n \leq L + k(L)} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right) \sim \frac{c k(L)}{g(L)} \sim k(L)\] which by our argument in the previous paragraph (\(c = g(L)\) functions as an upper bound, and \(c = g \left( L + k(L) \right)\) functions as a lower bound), implies that \[\sum_{L < n \leq L + k(L)} \frac{1}{2 g(n)} \mathbbm{1} \left( x_n \in \left[ x-g(n), x + g(n) \right] \right) \sim k(L)\] which shows that \(\rho_N (x) \to 1\) uniformly in \(x\), as required. All we have left is to compute \[\frac{1}{2 g(L)} \sum_{L < n \leq L + k(L)} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right) = \frac{1}{2 g(L)} \left( \sum_{1 \leq n \leq L + k(L)} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right)-\sum_{1 \leq n \leq L} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right) \right)\] By the definition of discrepancy, \[\abs{\sum_{1 \leq n \leq L + k(L)} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right)-2 c \left( L + k(L) \right)} \leq \left( L + k(L) \right) D_{L + k(L)}\] and \[\abs{\sum_{1 \leq n \leq L} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right)-2 c L} \leq L D_{L}\] and so \[\abs{\frac{1}{2 g(L)} \sum_{L < n \leq L + k(L)} \mathbbm{1} \left( x_n \in \left[ x-c, x + c \right] \right)-2 c k(L)} \leq \frac{L D_L + \left( L + k(L) \right) D_{L + k(L)}}{g(L)}\] In order for our sum to be asymptotically equal to \(2 c k(L)\), we need to have \[\lim_{L \to \infty} \frac{L D_L}{g(L) k(L)} = 0\] Choosing \(k(L) = f(L) \frac{d_L}{g(L)}\) where \(f(L)\) is some slowly growing function, by our regularity condition on \(g\) we get the desired result. The proof that \(h_{s, N} (x) \to 2 s\) is very similar. Notice that \[\frac{2 s}{N} \sum_{1 \leq n \leq N} \frac{1}{2 g(n)} \mathbbm{1} \left( x_n \in \left[ x-g(n) + \frac{s}{N}, x + g(n)-\frac{s}{N} \right] \right) \leq h_{s, N} (x) \leq\] \[\leq \frac{2 s}{N} \sum_{1 \leq n \leq N} \frac{1}{2 g(n)} \mathbbm{1} \left( x_n \in \left[ x-g(n)-\frac{s}{N}, x + g(n) + \frac{s}{N} \right] \right)\] where the lower bound follows from the fact that if \(x_n \in \left[ x-g(n) + \frac{s}{N}, x + g(n)-\frac{s}{N} \right]\) then \(\mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_n + z_n-x \right) \right] = \frac{s}{N g(n)}\), and the upper bound follows from the fact that if \(\mathbbm{1}_{s, N} \left( x-x_n-z_n \right)\) is positive with a non zero probability, then \(\abs{x_n-x} \leq g(n) + \frac{s}{N}\). Now, following exactly the same proof as before with \(g(n) \pm \frac{s}{N}\) instead of \(g(n)\) we get the desired result. Note that this is the point in the proof where we use the assumption \(N g(N) \to \infty\): this is precisely saying that the \(\pm \frac{s}{N}\) is negligible. Now, we bound the variance: \[\mathrm{Var} \left( X_N \right) = \mathbb{E} \left[ X_{N}^2 \right]-\mathbb{E} \left[ X_N \right]^2 =\] \[= \frac{1}{N^2} \sum_{\substack{1 \leq m_1, m_2, n_1, n_2 \leq N \\ m_1 \neq n_1, \ m_2 \neq n_2}} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_{n_1} + z_{n_1}-x_{m_1}-z_{m_1} \right) \mathbbm{1}_{s, N} \left( x_{n_2} + z_{n_2}-x_{m_2}-z_{m_2} \right) \right]-\mathbb{E} \left[ X_N \right]^2\] Notice that if \(m_1, n_1, m_2, n_2\) are distinct, the random variables \(z_{n_1}-z_{m_1}\) and \(z_{n_2}-z_{m_2}\) are independent, and therefore \[\frac{1}{N^2} \sum_{\substack{1 \leq m_1, m_2, n_1, n_2 \leq N \\ m_1, m_2, n_1, n_2 \ \mathrm{distinct}}} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_{n_1} + z_{n_1}-x_{m_1}-z_{m_1} \right) \mathbbm{1}_{s, N} \left( x_{n_2} + z_{n_2}-x_{m_2}-z_{m_2} \right) \right] =\] \[= \frac{1}{N^2} \sum_{\substack{1 \leq m_1, m_2, n_1, n_2 \leq N \\ m_1, m_2, n_1, n_2 \ \mathrm{distinct}}} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_{n_1} + z_{n_1}-x_{m_1}-z_{m_1} \right) \right] \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_{n_2} + z_{n_2}-x_{m_2}-z_{m_2} \right) \right] \leq \mathbb{E} \left[ X_N \right]^2\] which means that \[\mathrm{Var} \left( X_N \right) \leq \frac{1}{N^2} \sum_{1 \leq m_1, n_1, m_2, n_2 \leq N}' \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_{n_1} + z_{n_1}-x_{m_1}-z_{m_1} \right) \mathbbm{1}_{s, N} \left( x_{n_2} + z_{n_2}-x_{m_2}-z_{m_2} \right) \right]\] where \(\sum'\) means that the summation is taken over quadruples \(m_1, n_1, m_2, n_2\) such that \(m_1 \neq n_1, \ m_2 \neq n_2\) and the sets \(\left\{ m_1, n_1 \right\}, \ \left\{ m_2, n_2 \right\}\) have a nonempty intersection. If the sets have \(2\) elements in common, then \(m_1 = m_2, \ n_1 = n_2\) or \(m_1 = n_2, \ m_2 = n_1\), and so the sum over those quadruples is simply \[\frac{2 \mathbb{E} \left[ X_N \right]}{N}\] If the sets have an intersection of size \(1\), then the sum is \[\frac{4}{N^2} \sum_{\substack{1 \leq m, n_1, n_2 \leq N \\ m \neq n_1 \neq n_2 \neq m}} \mathbb{E} \left[ \mathbbm{1}_{s, N} \left( x_{n_1} + z_{n_1}-x_m-z_m \right) \mathbbm{1}_{s, N} \left( x_{n_2} + z_{n_2}-x_m-z_m \right) \right]\] Using the identity \[\frac{1}{N} \sum_{m = 1}^{N} \mathbb{E} \left[ f \left( x_m + z_m \right) \right] = \intop_{0}^{1} f(x) \rho_N (x) \mathrm{d} x\] a few times, we get that the variance is bounded by \[4 N \left( \intop_{0}^{1} \mathbbm{1}_{s, N} \left( x \right) \rho_N (x) \mathrm{d} x \right)^2 \left( \intop_{0}^{1} \rho_N (x) \mathrm{d} x \right) \sim \frac{16 s}{N}\] By Chebyshev's inequality we have \[\mathbb{P} \left( | X_N-\mathbb{E} \left[ X_N \right] | \geq \frac{1}{N^{1/4}} \right) \ll \frac{1}{N^{1/2}}\] and so by Borel-Cantelli, almost surely we have pair correlation along the subsequence \(X_{N^4}\). However, a standard argument shows that this is sufficient to get pair correlation, for example see. This concludes the proof of the theorem. Let us note an interesting fact about the proof: when computing the expectation, we showed that \(\rho_N (x)\) converges uniformly to \(1\), whereas at first glance it seems that we needed a much weaker result. However, as we have seen above, we usually have \(h_{s, N} (x) \sim 2 s \rho_N (x)\) and therefore \[\intop_{0}^{1} h_{s, N} (x) \rho_N (x) \mathrm{d} x \sim 2 s \intop_{0}^{1} \rho_N (x)^2 \mathrm{d} x\] This integral converges to \(2 s\) if and only if \[\intop_{0}^{1} \rho_N (x)^2 \mathrm{d} x \to 1\] Recall that \[\intop_{0}^{1} \rho_N (x) \mathrm{d} x = 1\] and so we have the required convergence if and only if \(\rho_N (x)\) converges to the constant \(1\) in \(L^2 \left[ 0, 1 \right]\). The variance bound is actually much more flexible than the expectation: all we need for the Chebyshev/Borel-Cantelli argument to work is \[\mathrm{Var} \left( X_N \right) \ll \frac{1}{N^{\varepsilon}}\] and in fact this can be further weakened to \[\mathrm{Var} \left( X_N \right) \ll \frac{1}{\left( \log N \right)^{1 + \varepsilon}}\] Therefore, if we can show that \(\rho_N (x)\) does not converge to \(1\) in \(L^2\), which is reasonable to expect if \(g(N)\) is smaller than the discrepancy, we can show that almost surely the pair correlation is too large to be Poissonian, and in particular we get that almost surely we do not have Poissonian pair correlation. In the next section, we shall partially follow this through for the sequence \(x_n = \alpha n\) for almost all \(\alpha\), but this is still far from a full converse. We think this is interesting, and a satisfactory converse theorem to Theorem [\[disc_pois\]](#disc_pois){reference-type="ref" reference="disc_pois"} would confirm this connection between discrepancy and pair correlation.
# Proof of Theorem [\[converse\]](#converse){reference-type="ref" reference="converse"}
We set up the same machinery as in the previous section. We show that almost surely, \[\limsup_{N \to \infty} X_{s, N} > 2 s\] which will imply lack of Poissonian pair correlation. Again, we do this via a second moment method. As in the previous section, we have \[\mathbb{E} \left[ X_N \right] = \intop_{0}^{1} h_{s, N} (x) \rho_N (x) \mathrm{d} x + o_N (1)\] As we have sketched in the previous section, \(h_{s, N} \sim 2 s \rho_N (x)\) and so if we show that \(\rho_N\) does not converge to \(1\) in \(L^2\) we will have \[\limsup_{N \to \infty} \mathbb{E} \left[ X_N \right] > 2 s\] By Khintchine's theorem, for almost all \(\alpha\), there exist infinitely many coprime \(p, q\) such that \[\abs{\alpha-\frac{p}{q}} \leq \frac{1}{q^2 \log q \log \log q}\] Let \(N = q \left( \log q \right)^{1/2} \left( \log \log q \right)^{1/3}\). From now on in this section, we will work only with \(N\) along this specific subsequence. Then, for \(1 \leq n \leq N\) \[\norm{n \alpha-n \frac{p}{q}} \leq \frac{N}{N^2 \left( \log \log q \right)^{1/3}} = o \left( \frac{1}{N} \right)\] and therefore in studying \(X_N\) we can replace the points \(n \alpha\) with \(n \frac{p}{q}\), as local statistics at scale \(\frac{1}{N}\) are unchanged. Assume for contradiction that \(\rho_N (x)\) converges to \(1\) in \(L^2\), and that \(\rho_{N / 2}\) converges to \(1\) in \(L^2\) as well. Then, so does \(2 \rho_N (x)-\rho_{N / 2} (x)\), which means that \[\frac{2}{N} \sum_{n = N / 2}^{N} \frac{1}{2 g(n)} \mathbbm{1} \left( x \in \left[ n \frac{p}{q}-g(n), n \frac{p}{q} + g(n) \right] \right)\] converges to \(1\) in \(L^2\). However, notice that the measure of the support of this function tends to \(0\), because its support is a union of intervals centered around at most \(q \leq \frac{N}{\left( \log N \right)^{1/2} \left( \log \log N \right)^{1/4}}\) distinct points, and each distinct point contributes to the set an interval of measure at most \(\frac{\left( \log N \right)^{1/2}}{N}\). Therefore, \(2 \rho_N-\rho_{N / 2}\) converges in measure to \(0\), and in particular does not converge to \(1\) in \(L^2\). Now we bound the variance. As in the previous section, it is bounded by \[4 N \left( \intop_{0}^{1} \mathbbm{1}_{s, N} \left( x \right) \rho_N (x) \mathrm{d} x \right)^2\] We will now give a slightly coarse pointwise bound on \(\rho_N (x)\): we show that uniformly in \(0 \leq x \leq 1\) we have \[\rho_N (x) \ll \log N\] We do this by splitting the sum into intervals of the form \(L \leq n \leq 2 L\). We now need to bound \[\frac{1}{g (2 L)} \sum_{n = 1}^{L} \mathbb{1} \left( x \in \left[ n \frac{p}{q}-g \left( L \right), n \frac{p}{q} + g \left( L \right) \right] \right)\] Clearly, this is at most \[\frac{g \left( L \right)}{g \left( 2 L \right)} \left( 1 + \frac{L}{q} \right) \ll 1 + \frac{L}{q}\] Summing these \(\log N\) intervals we get the desired result. This implies that we have \[\mathrm{Var} \left( X_N \right) \ll \frac{\left( \log N \right)^2}{N}\] Now applying Chebyshev's inequality and Borel-Cantelli along a subsequence shows that almost surely, \[\limsup_{N \to \infty} X_{s, N} > 2 s\] as required.
# Sketch of proof of Theorem [\[disc-k\]](#disc-k){reference-type="ref" reference="disc-k"}
It is enough to show that for each \(k\) individually, almost surely we have \(k\)-level Poissonian correlation, because a countable intersection of sets of full measure is a set of full measure. Define \[X_{s, N} = \frac{1}{N} \sum_{a \in \mathcal{X}_k} \mathbbm{1}_{s, N} \left( x_{a_1}-x_{a_2}, x_{a_1}-x_{a_3}, \dots, x_{a_1}-x_{a_k} \right)\] As before, we have Poissonian pair correlation if and only if \[\lim_{N \to \infty} X_{s, N} = \left( s_2-s_1 \right) \left( s_4-s_3 \right) \cdots \left( s_{2 k-2}-s_{2 k-3} \right)\] It is sufficient to prove this for each \(s\) individually, and once again we use the second moment method. The expectation now ends up being asymptotically equivalent to \[\left( s_2-s_1 \right) \left( s_4-s_3 \right) \cdots \left( s_{2 k-2}-s_{2 k-3} \right) \intop_{0}^{1} \rho_N (x)^k \mathrm{d} x\] and and the asymptotic still follows from the fact that \(\rho_N\) uniformly converges to \(1\). As for the variance, the way to take care of it is to split \(\mathbb{E} \left[ X_{s, N}^2 \right]\) into sums of the form \[\sum_{\substack{a, b \in \mathcal{X}_k \\ \abs{a \cap b} = \ell}} \cdots\] that is, the summation is over pairs of sequences \(a, b\) with exactly \(\ell\) common elements, exactly like what we did above. The contribution of the sum for \(\ell = 0\) is at most \(\mathbb{E} \left[ X_{s, N} \right]^2\) in the same manner as we did for pair correlation, and the contribution for each \(\ell \geq 1\) is easily seen to be (by the same method) \(\ll \frac{1}{N}\), and using Chebyshev's inequality and Borel-Cantelli proves the theorem. | {'timestamp': '2021-09-14T02:17:13', 'yymm': '2109', 'arxiv_id': '2109.05379', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05379'} |
null | null |
null | null |
null | null |
null | null |
null | null |
# [\[sec:1\]]{#sec:1 label="sec:1"}Introduction
States which possess nonclassical features are an important resources for quantum information processing and the investigation of fundamental problems in quantum theory. It has been shown that squeezed states of radiation fields has been can be considered truly quantum. In recent years studies concerning squeezing especially quadrature squeezing of radiation fields has seen considerable attention as it may have applcaition in optical communication and information theory, gravitatioanl wave detection, quantum teleportation , dense coding, resonance fluorscence, and quantum cryptography. Furthermore, with the rapid development of the techniques for making higher-order correlation measurements in quantum optics and laser physics, the high-order squeezing effects of radiation fields have also became a hot topic in state optimization researches. Higher-order squeezing of radiation fields was first introduced by Hong and Manel in 1985, and Hilley defined another type higher-order squeezing, named amplitude-squared squeezing (ASS) of the electromagnetic field in 1987.[ ]{style="color: brown"}Following this work the highe-squeezing of radiation fields has been investigated across many fields of research.. Squeezing is an inherent feature of nonclassical states, and its improvement requires optimization. Some states do not initially possess squeezing, but after undergoing an optimization process, they may possess a pronounced squeezing effect, The single-photon-added coherent state (SPACS) is a typical example. SPACAS are created by adding the creation operator \(a^{\dagger}\) to the coherent state, and this optimization changes the coherent state from semi-classical to a new quantum state which possess squeezing. Since this state has wide application across many quantum information processes including quantum communication , quantum key distribution , and quantum digital signature, the optimization for this state is worthy of study, in particular, it may provide new methods to the implementations related processes. On the other hand, the weak signal amplification technique proposed in 1988[ ]{style="color: brown"} by Aharonov, Albert, and Vaidman is widely used in state optimization and precision measurement problems. Most recently, one of the authors of this paper investigated the effects of postselected von Neumann measurement on the properties of single-mode radiation fields and found that postselected von Neumann measurement changed the photon statistics and quadrature squeezing of radiation fields for different anomalous weak values and coupling strengths. However, to the best of our knowledge, the effects of postselected von Neumann measurement on higher-order squeezing and phase-space distribution of SPACS have not been previously investigated. In this work, motivated by our prior work, we study the squeezing and Wigner function of SPACS after postselected von Neumann measurement. In this work, we take the spatial and polarization degrees of freedom of SPACS as a measuring device (pointer) and system, respectively, and consider all orders of the time evolution operator. Following determination of the final state of the pointer, we check the criteria for existence of squeezing of SPACS, and found that the postselected measurement has positve effects on squeezing of SPACS in the weak measurement regime. Furthermore, we investigate the state-distance and the Wigner function of the SPACS after measurement. We found that with increasing coupling strength, the original SPACS spoiled significantly, and the state exhibited more pronounced negative areas as well as interference structures in phase space after postselected measurement.[ ]{style="color: brown"}We observed that the postselected von Neumann measurement has positive effects on its nonclassicality including squeezing effects especially in the weak measurement regime. These results can be considered a result of weak value amplification of the weak measurement technique. This paper is organized as follows. In Sec. [\[sec:2\]](#sec:2){reference-type="ref" reference="sec:2"}, we introduce the main concepts of our scheme and derive the final pointer state after postselected measurement which will be used throughout the study. In Sec. [\[sec:3\]](#sec:3){reference-type="ref" reference="sec:3"}, we give the details of ordinary squeezing and ASS effects of the final pointer state. In Sec. [\[sec:4\]](#sec:4){reference-type="ref" reference="sec:4"}, we investigate the state distance and the Wigner function SPACS after measurement. A conclusion is given in Sec. [\[sec:5\]](#sec:5){reference-type="ref" reference="sec:5"}.
# [\[sec:2\]]{#sec:2 label="sec:2"}Model and theory
In this section, we introduce the basic concepts of postselected von Neumann measurement and give the expression of the final pointer state which we use in this paper. We know that every measurement problems consists of three main parts including a pointer(measuring device), measuring system and the environment. In the current work, we take the spatial and polarization degrees of freedom of SPACS as the pointer and system, respectively. In general, in measurement problems we want to determine the system information of interest by comparing the state-shifts of the pointer after measurement finishes, and we do not consider spoiling of the pointer in the entire measurement process. Here, contrary to the standard goal of the measurement, we investigate the effects of pre-and post-selected measurement taken on a beam's polarization(measured system) on the inherent properties of a beam's spatial component (pointer). In the measurement process, the system and pointer Hamiltonians doesn't effect the final read outs, so it is sufficient to only consider their interaction Hamiltonian for our purposes. According to standard von Neumann measurement theory, the interaction Hamiltonian between the system and the pointer takes the form \[\hat{H}=g(t)\hat{A}\otimes\hat{P}.\label{eq:2}\] Here, \(\hat{A}\) is the system observable we want to measure, and \(\hat{P}\) is the momentum operator of the pointer conjugated with the position operator, \(\left[\hat{X},\hat{P}\right]=i\). \(g(t)\) is the coupling strength function between the system and pointer and it is assumed exponentially small except during a period of interaction time of order \(T\), and is normalized according to \(\int_{-\infty}^{+\infty}g(t)dt=\int_{0}^{T}g(t)dt=g_{0}\). In this work, we assume that the system observable \(A\) is Pauli \(x\) matrix, i.e., \[\hat{A}=\hat{\sigma}_{x}=\vert H\rangle\langle V\vert+\vert V\rangle\langle H\vert=\left(\begin{array}{cc} 0 & \ \ 1\\ 1 & 0 \end{array}\right)\label{eq:1}\] Here, \(\vert H\rangle\equiv(1,0)^{T}\) and \(\vert V\rangle\equiv(0,1)^{T}\)represent the horizontal and vertical polarization of the beam, respectively. We also assume that in our scheme the pointer and measurement system are initially prepared to \[\vert\phi\rangle=\gamma a^{\dagger}\vert\alpha\rangle,\ \ \ \ \gamma=\frac{1}{\sqrt{1+\vert\alpha\vert^{2}}}\label{eq:3-2}\] and \[\vert\psi_{i}\rangle=\cos\frac{\varphi}{2}\vert H\rangle+e^{i\delta}\sin\frac{\varphi}{2}\vert V\rangle,\label{eq:3-1-1}\] respectively. Here, \(\alpha=re^{i\theta}\) and \(\delta\in[0,2\pi]\) and \(\varphi\in[0,\pi)\). Here, we are reminded that in weak measurement theory, the interaction strength between the system and measurement is weak, and it is enough to only consider the evolution of the unitary operator up to its first order. However, if we want to connect the weak and strong measurement and investigate the measurement feedback of postselected weak measurement, and analyze experimental results obtained in non-ideal measurements, the full-order evolution of the unitary operator is needed, We call this kind of measurement a postselected von Neumann measurement. Thus, the evolution operator of this total system corresponding to the interaction Hamiltonian, Eq. ([\[eq:2\]](#eq:2){reference-type="ref" reference="eq:2"}), is evaluated as \[e^{-ig_{0}\sigma_{x}\otimes P}=\frac{1}{2}\left(\hat{I}+\hat{\sigma}_{x}\right)\otimes D\left(\frac{s}{2}\right)+\frac{1}{2}\left(\hat{I}-\hat{\sigma}_{x}\right)\otimes D\left(-\frac{s}{2}\right)\label{eq:7}\] since\(\hat{\sigma}_{x}^{2}=1\). Here,\(s=\frac{g_{0}}{\sigma}\) is the ratio between the coupling strength and beam width, and it can characterize the measurement types i.e. the measurement is considered a weak measurement (strong measurement) if \(s<1\) (\(s>1\)). \(D(\frac{s}{2})\) is the displacement operator defined as \(D(\alpha)=e^{\alpha\hat{a}^{\dagger}-\alpha^{\ast}\hat{a}}\). The results of our current research are valid for weak and strong measurement regimes since we take into account the all orders of the time evolution operator, Eq. ([\[eq:7\]](#eq:7){reference-type="ref" reference="eq:7"}). In the above calculation we use the definition of the momentum operator represented in Fock space in terms of an annihilation (creation) operator \(\hat{a}\) (\(\hat{a}^{\dagger}\)), i.g., \[\hat{P}=\frac{i}{2\sigma}\left(a^{\dagger}-a\right)\label{eq:P}\] where \(\sigma\) is the width of the beam. Thus, the total state of the system, \(\vert\psi_{i}\rangle\otimes\vert\phi\rangle\), after the time evolution becomes \[\begin{aligned}
\vert\Psi\rangle & =e^{-ig_{0}\sigma_{x}\otimes P}\vert\psi_{i}\rangle\otimes\vert\phi\rangle\nonumber \\ & =\frac{1}{2}\left[\left(\hat{I}+\hat{\sigma}_{x}\right)\otimes D\left(\frac{s}{2}\right)+\left(\hat{I}-\hat{\sigma}_{x}\!\right)\otimes D\left(\!\frac{-s}{2}\right)\right]\!\vert\psi_{i}\rangle\otimes\vert\phi\rangle\label{eq:8} \end{aligned}\] After we take a strong projective measurement of the polarization degree of the beam with posts-elected state \(\vert\psi_{f}\rangle=\vert H\rangle\), the above total system state gives us the final state of the pointer, and its normalized expression reads as \[\vert\Phi\rangle=\frac{\kappa}{\sqrt{2}}\left[\left(1+\langle\sigma_{x}\rangle_{w}\right)D\left(\frac{s}{2}\right)+\left(1-\langle\sigma_{x}\rangle_{w}\right)D\left(-\frac{s}{2}\right)\right]\vert\phi\rangle.\label{eq:10-2}\] Here, \[\begin{aligned}
\kappa^{-2} & =\!1+\!\vert\langle\sigma_{x}\rangle\vert^{2}\!+\gamma^{2}e^{-\frac{s^{2}}{2}}Re[(1+\langle\sigma_{x}\rangle_{w}^{\ast})(1-\!\langle\sigma_{x}\rangle_{w})\times\nonumber \\ & (\gamma^{-2}-\text{\ensuremath{s^{2}+\alpha s-\alpha^{\ast}s}})e^{2si\Im(\alpha)}]\label{eq:11-1} \end{aligned}\] is the normalization coefficient, and the weak value of the system observable \(\hat{\sigma}_{x}\) is given by \[\langle\sigma_{x}\rangle_{w}=\frac{\langle\psi_{f}\vert\sigma_{x}\vert\psi_{i}\rangle}{\langle\psi_{f}\vert\psi_{i}\rangle}=e^{i\delta}\tan\frac{\varphi}{2}.\label{eq:3}\] In general, the expectation value of \(\sigma_{x}\) is bounded \(-1\le\langle\sigma_{x}\rangle\le1\) for any associated system state. However, as we see in Eq. ([\[eq:3\]](#eq:3){reference-type="ref" reference="eq:3"}), the weak values of the observable \(\sigma_{x}\) can take arbitrary large numbers with small successful post-selection probability \(P_{s}=\vert\langle\psi_{f}\vert\psi_{i}\rangle\vert^{2}=\cos^{2}\frac{\varphi}{2}\). This weak value feature is used to amplify very weak but useful information on various of related physical systems. The state given in Eq. ( [\[eq:10-2\]](#eq:10-2){reference-type="ref" reference="eq:10-2"}) is a spoiled version of SPACS after postselected measurement. In the next sections, we study squeezing effects, and nonclassicality features characterized by the Wigner function.
# [\[sec:3\]]{#sec:3 label="sec:3"}Ordinary and amplitude square squeezing
In this section, we check the ordinary (first-order) and ASS (second order) squeezing effects of SPACS after postselected von Neumann measurement.The squeezing effect is one of the non-classical phenomena unique to the quantum light field. The squeezing reflects the non-classical statistical properties of the optical field by a noise component lower than that of the coherent state. In other words, the noise of an orthogonal component of the squeezed light is lower than the noise of the corresponding component of the coherent state light field. In practice, if this component is used to transmit information, a higher signal-to-noise ratio can be obtained than that of the coherent state. Consider a single mode of electromagnetic field of frequency \(\text{\ensuremath{\omega}}\) with creation and annihilation operator \(a^{\dagger},\[a\). The quadrature and square of the field mode amplitude can be defined by operators \(X_{\theta}\) and \(Y_{\theta}\) as \]X_{\theta}\equiv\frac{1}{2}\left(ae^{-i\theta}+a^{\dagger}e^{i\theta}\right)\label{eq:11}\[ and \]Y_{\theta}\equiv\frac{1}{2}\left(a^{2}e^{-i\theta}+a^{\dagger2}e^{i\theta}\right),\label{eq:12}\[ respectively. For these operators, if \(\triangle X_{\theta}\equiv X_{\theta}-\langle X_{\theta}\rangle\), \(\triangle Y_{\theta}\equiv Y_{\theta}-\langle Y_{\theta}\rangle\), the minimum variances are \]\langle(\triangle X_{\theta})^{2}\rangle_{min}=\frac{1}{4}+\frac{1}{2}\left[\left(\langle a^{\dagger}a\rangle-\vert\langle a\rangle\vert^{2}\right)-\vert\langle a^{2}\rangle-\langle a\rangle^{2}\vert\right]\label{eq:13}\[ \]\begin{aligned}
\langle(\triangle Y_{\theta})^{2}\rangle_{min} & =\langle a^{\dagger}a+\frac{1}{2}\rangle\label{eq:14}\\ & +\frac{1}{2}\left[\langle a^{\dagger2}a^{2}\rangle-\vert\langle a^{2}\rangle\vert^{2}-\vert\langle a^{4}\rangle-\langle a^{2}\rangle^{2}\vert\right]\nonumber \end{aligned}\[ where \(a\) and \(a^{\dagger}\) are annihilation and creation operators of the radiation field. If \(\langle(\triangle X_{\theta})^{2}\rangle_{min}<\frac{1}{4}\), \(X_{\theta}\) is said to be ordinary squeezed and if \(\langle(\triangle Y_{\theta})^{2}\rangle_{min}<\langle a^{\dagger}a+\frac{1}{2}\rangle\), \(Y_{\theta}\) is said to be ASS. These conditions can be rewritten as \]S_{os}=\langle a^{\dagger}a\rangle-\vert\langle a\rangle\vert^{2}-\vert\langle a^{2}\rangle-\langle a\rangle^{2}\vert<0\label{eq:15}\[ \]S_{ass}=\langle a^{\dagger2}a^{2}\rangle-\vert\langle a^{2}\rangle\vert^{2}-\vert\langle a^{4}\rangle-\langle a^{2}\rangle^{2}\vert<0.\label{eq:16}\[ Thus, the system characterized by any wave function may exhibit non-classical features if it satisfies Eqs. ([\[eq:15\]](#eq:15){reference-type="ref" reference="eq:15"}-[\[eq:16\]](#eq:16){reference-type="ref" reference="eq:16"}). To achieve our goal, we first have to calculate the above related quantities and their explicit expressions under the state \(\vert\Phi\rangle\).these are listed below.
Using the expression for \(S_{os}\), the curves for this quantity are plotted, and the analytical results are shown in Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}. In Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}(a), we fixed the parameter \(r=1\) and plot the \(S_{os}\) as a function of coupling strength \(s\) for different weak values quantified by \(\varphi\). As we observed, when there is no interaction between system and poiner (\(s=0\)), there is no ordinary squeezing effect of initial SPACS at the \(r=1\) point. However, in moderate coupling strength regions such as \(0<s\)\<2, the ordinary squeezing effect of SPACS is proportional to the weak value, i.e. the larger the weak value, the better its squeezing effect. From Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}(a) we also can see that the ordinary squeezing effect of the light field gradually disappears and tends to the same value for different weak values with increasing coupling strength \(s\) in the strong measurement regime. In Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}(b), we plot \(S_{os}\) as a function of the state parameter \(r\) in the weak measurement regime by fixing the coupling strength \(s\), i.g, \(s=0.5\). It is very clear from the curves presented in Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"} (b) that the ordinary squeezing effect of SPACS is increased when increasing the weak value, especially when \(\varphi\) is taken as \(\frac{7\pi}{9}\). Furthermore, along with the increasing \(r\) (for \(r\) values exceeding 1.5), the squeezing effect of the field for different weak values tended to be the same. According to von Neumann measurement theory, when the interaction strength is too large, the system is strongly measured and the size of the weak value has little impact on the squeezing effect.This statement can also be observed in Fig. [\[Fig1\]](#Fig1){reference-type="ref" reference="Fig1"}(a) and (b). In the weak measurement regime the SPACS showed a good ordinary squeezing effect after postselected measurement with large weak values, and this can be seen as a result of the signal amplification feature of the weak measurement technique. The quantity \(S_{ass}\)can characterize the ASS of SPACS if it takes negative values, and in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"} it is plotted as a function of various system parameters. As indicated in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"} (a), when we fixed the coherent state parameter \(r\) ,the \(S_{ass}\) can take negative values in the weak measurement regime (\(s<1\)) and its negativity increases when increasing the weak value quantified by \(\varphi\). That is to say, in the weak measurement regime, the magnitude of the weak value has a linear relationship with the ASS effect of SPACS, i.g. The larger the weak value, the better the ASS effect. However, by increasing the coupling strength, the value of \(S_{ass}\) became larger than zero and it indicates that there is no ASS effect on SPACS in the postselected strong measurement regime (\(s>1\)) no matter how large the value is taken. In order to further investigate the ASS of the radiation field in the weak measurement regime, we plot the \(S_{ass}\) as a function of the coherent state parameter \(r\) for different weak values with \(s=0.5\). the analytical results are shown in Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}b. We can see that when \(r\) is relatively small, there is an ASS effect no matter how large the weak value becomes. By increasing the system parameter \(r\), \(S_{ass}\) takes negative values and its negativity is proportional to \(r\). From Fig. [\[Fig2\]](#Fig2){reference-type="ref" reference="Fig2"}a we can also observe that in the weak measurement regime, the weak values have positive effects on the ASS of SPACS, and it can also be considered a result of the weak signal amplification feature of the postselected weak measurement technique.
# [\[sec:4\]]{#sec:4 label="sec:4"}State distance and Wigner function
The postselected measurement taken on polarization degree of freedom of the beam could spoil the inherent properties presented in its spatial part. Before we investigate the phase-space distribution of SPACS after postselected von Neumanm measurement, we check the similarity between the initial SPACS \(\vert\phi\rangle\) and the state \(\vert\Phi\rangle\) after measurement. The state distance between those two states can be evaluated by \]F=\vert\langle\phi\vert\Phi\rangle\vert^{2},\label{eq:23}\[ and its value is bounded \(0\le F\le1\). If \(F=1\) (\(F=0\)), then the two states are totally same (totally different). The \(F\) in our case can be calculated after substituting equations Eq. ([\[eq:3-2\]](#eq:3-2){reference-type="ref" reference="eq:3-2"})and Eq. ([\[eq:10-2\]](#eq:10-2){reference-type="ref" reference="eq:10-2"}) into the Eq. ([\[eq:23\]](#eq:23){reference-type="ref" reference="eq:23"}), and the analytical results are shown in Fig. [\[Fig:3\]](#Fig:3){reference-type="ref" reference="Fig:3"}. In Fig. [\[Fig:3\]](#Fig:3){reference-type="ref" reference="Fig:3"} we present the state distance \(F\) as a function of system parameter \(r\) for different coupling strengths with a fixed large weak value. As shown in Fig. [\[Fig:3\]](#Fig:3){reference-type="ref" reference="Fig:3"}, in the weak coupling regime (\(s=0.5\)), the state after the postselected measurement maintains similarity with the the coherent state parameter \(r\). However, with increasing the measurement strength, the initial state \(\vert\phi\rangle\) is spoiled and the similarity between the pointer states before and after the measurement is decreases. In order to further explain the squeezing effects of SPACS after postselected von Neumann measurement, in the rest of this section we study the Wigner function of \(\vert\Phi\rangle\). The Wigner distribution function is the closest quantum analogue of the classical distribution function in phase space. According to the value of the Wigner function we can intuitively determine the strength of its quantum nature, and the negative value of the Wigner function proves the nonclassicality of the quantum state. The Wigner function exists for any state, and it is defined as the two-dimensional Fourier transform of the symmetric order characteristic function. Thus, the Wigner function for the state \(\rho=\vert\Phi\rangle\langle\Phi\vert\) is written as \]W(z)\equiv\frac{1}{\pi^{2}}\int_{-\infty}^{+\infty}\exp(\lambda^{\ast}z-\lambda z^{\ast})C_{W}(\lambda)d^{2}\lambda,\label{eq:35-1}\[ where \(C_{N}(\lambda)\) is the normal ordered characteristic function,and is defined as \]C_{W}(\lambda)=Tr\left[\rho e^{\lambda a^{\dagger}-\lambda^{\ast}a}\right].\label{eq:34}\[ Using the notation \(\lambda^{\prime},\]\lambda^{\prime\prime}\) for the real and imaginary parts of \(\lambda\) and setting \(z=x+ip\) to emphasize the analogy between the radiation field quadratures and the normalized dimensionless position and momentum observables of the beam in phase space. We can rewrite the definition of the Wigner function in terms of \(x,p\) and \(\lambda^{\prime},\lambda^{\prime\prime}\) as \[W(x,p)=\frac{1}{\pi^{2}}\int_{-\infty}^{+\infty}e^{2i(p\lambda^{\prime}-x\lambda^{\prime\prime})}C_{W}(\lambda)d\lambda^{\prime}d\lambda^{\prime\prime}.\label{eq:26}\] By substituting the final normalized pointer state \(\vert\Phi\rangle\) into Eq. ([\[eq:26\]](#eq:26){reference-type="ref" reference="eq:26"}), we can calculate the explicit expression of its Wigner function and it reads as \[\begin{aligned}
W(z) & =\frac{2\vert\kappa\vert^{2}}{\pi(1+\vert\alpha\vert^{2})}e^{-2\vert z-\alpha\vert^{2}}\times\nonumber \\ & \{\vert1+\langle\sigma_{x}\rangle_{w}\vert^{2}w(\Gamma)+\vert1-\langle\sigma_{x}\rangle_{w}\vert^{2}w(-\Gamma)\nonumber \\ & \!\!\!\!\!\!\!\!\!+2\left(-1+\vert2z-\alpha\vert^{2}\right)Re[(1+\langle\sigma_{x}\rangle_{w})^{\ast}(1-\langle\sigma_{x}\rangle_{w})e^{2isIm[z]}]\}.\label{eq:35} \end{aligned}\] with \[\begin{aligned}
w(\Gamma) & =e^{-\frac{1}{2}s^{2}}e^{-2(Re[\alpha]-Re[z])s}\times\nonumber \\ & \left(-1+\vert2z-\alpha\vert^{2}+2s(Re[\alpha]-2Re[z]+\frac{s}{2})\right) \end{aligned}\] This is a real Wigner function and its value is bounded \(-\frac{2}{\pi}\leq W(\alpha)\leq\frac{2}{\pi}\) in whole phase space. To depict the effects of the postselected von Neumann measurement on the non-classical feature of SPACS, in Fig. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"} we plot its curves for different parametric coherent state parameters \(r\) and coupling strengths \(s\). Each column from left to right in-turn indicate the different coherent state parameters \(r\) for 0, 1 and 2, and each row from up to down represent the different coupling strengths \(s\) for \(0,0.5\) and \(2\). It is observed that the positive peak of the Wigner function moves from the center to the edge position in phase space and its shape gradually becomes irregular with changing coupling strength \(s\). From the first row (see Figs. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"}a-c ) we can see that the original SPACS exhibit inherent features changing from single photon state to coherent states with gradually increasing coherent state parameter \(r\). Figs. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"}d-k indicate the phase space density function \(W\left(z\right)\) after postselected von Neumann measurement. Fig. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"} d-f represent the Wigner function for fixed weak interaction strength \(s=0.5.\) It can be observed that the Wigner function distribution shows squeezing in phase space compared to the original SPACS and this kind of squeezing is pronounced with increasing coupling strength (see Figs. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"}g-k ). Furthermore, in Figs. [\[Fig:4\]](#Fig:4){reference-type="ref" reference="Fig:4"}(g-k) we can see that in the strong measurement regime significant interference structures manifest and the negative regions become larger than the initial pointer state. As mentioned above, the existence of and progressively stronger negative regions of the Wigner function in phase space indicates the degree of nonclassicality of the associated state. From the above analysis we can conclude that after the postselected von Neumann measurement, the phase space distribution of SPACS is not only squeezed but the nonclassicality is also pronounced in the strong measurement regime.
# [\[sec:5\]]{#sec:5 label="sec:5"}Conclusion
In this paper we have studied the squeezing and Wigner function of SPACS after postselected von Neumann measurement. In order to achieve our goal, we first determined the final state of the pointer state along with the standard measurement process. We examined the ordinary (first-order) and ASS effects after measurement, and found that in the weak measurement region, the ordinary squeezing and ASS of the light field increased significantly as the weak value increased. To further explain our work, we examined the similarity between the initial SPACS and the state after measurement. We observed that under weak coupling, the state after the postselected measurement maintains similarity with the initial state. However, as the intensity of the measurement increases, the similarity between them gradually decreased and indicated that the measurement spoils the system state if the measurement is strong. We also investigate the Wigner function of the system after postselected measurement. It is observed that following the postselected von Neumann measurement, the phase space distribution of SPACS is not only squeezed, but also adevelops significant interference structures in the strongly measured regime. It also possess pronounced nonclassicality characterized with a large negative area in phase space. We anticipate that the theoretical scheme in this paper may provide an effective method for solving practical problems in quantum information processing associated with SPACS. | {'timestamp': '2021-09-15T02:13:47', 'yymm': '2109', 'arxiv_id': '2109.05423', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05423'} |
null | null |
# Introduction {#sec:intro}
Advancements in deep learning have brought performance improvements in acoustic and language modeling, yielding robust automatic speech recognition (ASR) systems in many languages. However, such systems require a large amount of speech data and the associated transcriptions. It is tough to collect large volumes of paired speech and transcriptions for most low-resource languages. It is estimated that only about 1% of the world languages have the minimum amount of data that is needed to train an ASR. However, in many cases, especially for languages in south Asia, we can find a "close enough" language with the same set (or a superset) of phonemes as the low-resource language and enough resources for building an ASR. In this work, we show that a better performing ASR can be built for the low-resource language using unsupervised domain adaptation (UDA) of acoustic models from the corresponding high-resource language. This method has the benefit of modeling on real data in comparison to the data augmentation techniques like vocal tract length perturbation, speech and tempo perturbation, noise addition, data synthesis, and spectral augmentation, where the modeling makes use of the synthetic data as well. Unsupervised domain adaptation (UDA) has been successfully applied to various tasks to alleviate the shift between the train and test distributions. shows good adaptation performance in the classification task on digit image datasets having considerable domain shifts. They learn features that are discriminative for the image classification task and invariant to the domain. They introduce a gradient reversal layer (GRL) for achieving this objective. reports performance improvements in speech recognition for data shifted in the domain by gender and accent. employs GRL layers to reduce the mismatch between train and test domains in the task of emotion recognition from speech data. uses the GRL approach to improve the word error rate (WER) in speech recognition with the source domain data as clean speech and the target domain data as contaminated speech. They also show the robustness of the approach to the domain shifts caused by the differences in datasets. The basic UDA scheme with a GRL tries to learn domain invariant features but ignores the individual characteristics of each domain. introduces domain separation networks (DSN) and shows improvements in a range of UDA scenarios in the image classification task. They learn two representations: one specific to each domain and the other common to both domains. uses DSN for adaptation from clean speech to noisy speech. In this work, we explore the feasibility of the UDA schemes in the ASR task on low-resource languages that share the same acoustic space with a high-resource language. Specifically, we place the following assumptions in the selection of high and low resource language pairs.
1. The acoustic space spanned by the low-resource language is a subspace of that of the high-resource language. This requires the phoneme set of the low-resource language to be a subset of the high-resource language.
2. There exists a high-resource language with reasonable amount of paired audio data and transcriptions to build an ASR.
3. The low-resource language has enough text data available to train the language models.
4. The speech data available in the low-resource language is quite limited, and the transcriptions are not available.
Though the above assumptions seem quite constrictive, we can easily find a few low-resource languages in the Indian subcontinent which share a common acoustic space with a reasonably high-resource language. In this work, we use Hindi as the high-resource language and Sanskrit as the low-resource language, both belonging to the Indo-Aryan language family. Hindi is written in the Devanagari script, and many of its words are derived from Sanskrit. However, there exist substantial differences in the vocabulary and pronunciation between the two languages. One of the important differences is that the schwa, implicit in each consonant of the script, is not pronounced at the end of words and in some other contexts in Hindi. The script does not tell us when the schwa should be deleted, a phenomenon known as schwa deletion. For example, the word for \"salty\" is pronounced as *nam'kīn* in Hindi and not *namakīna*. Another difference is the pitch accents that are common in Sanskrit. Despite the above differences, these languages share a common phoneme set. So we can intuitively argue that there exists a domain shift between the distributions of acoustic features in Hindi and Sanskrit. This makes us believe that the speech recognition problem in Sanskrit, an extremely low-resource language, may be posed as an UDA problem from Hindi, a language with a reasonably good collection of annotated audio.
# Unsupervised domain adaptation for acoustic modeling {#sec:uda}
We pose the problem of acoustic modeling in Sanskrit as an unsupervised domain adaptation task from Hindi. We build a deep neural network (DNN)-hidden Markov model (HMM) ASR system for Sanskrit with the domain-independent acoustic models learned from Hindi through UDA approaches. We make use of the UDA schemes introduced in and.
## Adversarial training using GRL {#sec:grl}
In GRL-based adversarial training, we try to learn a feature representation invariant to the domain, but good enough in discriminating the senone labels. The neural network architecture for learning the acoustic model consists of three parts: feature extractor, senone classifier, and domain classifier. A block diagram of the UDA architecture employing GRL is shown in Fig. [\[fig:blockdiagram\]](#fig:blockdiagram){reference-type="ref" reference="fig:blockdiagram"}. Feature extractor \(G_f\) maps the input acoustic features \(\mathbf{x}\) to an internal representation \(\mathbf{f} \in \mathcal{R}^D\). Senone classifier \(G_y\) maps the output of the feature extractor to the senone labels \(\mathbf{y}\) whereas the domain classifier \(G_d\) maps them to domain labels \(\mathbf{d}\). We train the network to minimize the senone classification loss during the training phase by optimizing the parameters in the feature extractor \(G_f\) and senone classifier \(G_y\). This makes the network look for features that are capable of discriminating the senone labels. To make the features domain-invariant, we optimize the parameters of the feature extractor to maximize the domain classification loss. But at the same time, the parameters of the domain classifier \(G_d\) are optimized to minimize the domain classification loss. This is achieved by introducing a GRL between the feature extractor \(G_f\) and domain classifier \(G_d\). During the forward pass, GRL acts as an identity transform. GRL takes the gradient from the subsequent layer during the backward pass, multiplies it with \(-\alpha\), a hyperparameter to control the trade-off between senone discrimination and domain-invariance, and passes it to the preceding layer. During the inference time, the domain classifier and GRL are ignored. The acoustic feature vectors are passed through the feature extractor and the senone classifier, and senone labels are predicted.
## Domain separation networks {#sec:dsn}
A block diagram of the DSN architecture is shown in Fig. [\[fig:dsn\]](#fig:dsn){reference-type="ref" reference="fig:dsn"}. They model both the private and shared components of the domain representation. Private encoders \(E_p^t(.)\) and \(E_p^s(.)\) extract components \(f_p^t\) and \(f_p^s\) which are specific to the target and source domains. Shared encoder \(E_c(.)\) is common to both domains and extract the shared components \(f_c^t\) and \(f_c^s\). Shared decoder \(D\) tries to reconstruct the input using the private and shared components. \(G(.)\), the senone classifier maps \(f_c^s\) to the senone label \(\mathbf{\hat{y}}\). The domain classifier \(Z(.)\) maps the shared components \(f_c^s\) and \(f_c^t\) to their respective domain labels \(\mathbf{\hat{d}}\). The network is trained to minimize the following loss function with respect to the parameters of \(E_p\), \(E_c\), \(G\) and \(Z\): \[L = L_{class} + {\beta} L_{sim} + {\gamma} L_{diff} + {\delta} L_{recon}\] where \(\beta, \gamma\), and \(\delta\) are hyperparameters. \(L_{class}\) represents the senone classification loss and is applied only to the source domain. It is computed as the negative log-likelihood of the ground-truth senone labels. \(L_{sim}\) represents the domain adversarial similarity loss and is computed as the negative log-likelihood of the domain labels. \(L_{sim}\) ensures that the shared components \(f_c^t\) and \(f_c^s\) are as similar as possible irrespective of their domain so that the domain classifier cannot reliably predict the domain of the sample from its shared representation. Parameters of the domain classifier \(Z(.)\) are trained to minimize the domain classification loss while the parameters of the shared encoder \(E_c\) are trained to maximise the domain classification loss. This is also accomplished with a GRL. \(L_{diff}\) encourages the shared component \(f_c\) and the private component \(f_p\) to encode different aspects of the input. This is achieved by imposing a soft subspace orthogonality constraint between the private and shared components. \[L_{diff} = \| \mathbf{F}_c^{sT} \mathbf{F}_p^s \|^2_F + \| \mathbf{F}_c^{tT} \mathbf{F}_p^t \|^2_F\] where \(\mathbf{F}_c^{s}\), \(\mathbf{F}_c^{t}\), \(\mathbf{F}_p^s\) and \(\mathbf{F}_p^t\) are matrices with \(f_c^s\), \(f_c^t\), \(f_p^s\) and \(f_p^t\) as rows. \(\|.\|_F\) denotes the Frobenius norm. \(L_{recon}\) is the reconstruction loss, computed as the mean squared error (MSE) between \(\mathbf{x}\) and \(\hat{\mathbf{x}} = D (E_c(\mathbf{x})+E_p(\mathbf{x}))\). \[L_{recon} = \sum_{i=1}^{N_s} \| \mathbf{x}_i^s-\hat{\mathbf{x}}_i^s\|^2 + \sum_{i=1}^{N_t} \| \mathbf{x}_i^t-\hat{\mathbf{x}}_i^t\|^2\] where \(N_s\) and \(N_t\) represent the number of speech frames from the source and target domains. We also validate the performance of the system with scale-invariant mean squared error (SIMSE) which is computed as: \[L_{SIMSE} = \frac{1}{k} \| \mathbf{x}-\hat{\mathbf{x}} \|_2^2-\frac{1}{k^2} (\left[\mathbf{x}-\hat{\mathbf{x}}\right]. \mathbf{1}_k)^2\] where \(k\) in the dimension of the input vector \(\mathbf{x}\), \(\mathbf{1}_k\) is a \(k\)-dimensional vector of ones and \(\|.\|_2\) is the \(L_2\) norm.
# Experimental setup {#sec:setup}
## Datasets used for the study
We primarily use a Hindi dataset in the source domain. We also use a Telugu dataset in the ablation studies. Both these datasets are the same as the ones used in multilingual and code-switching ASR challenge, Interspeech-2021. The details of both these datasets are available in. Both the speech data and the corresponding transcriptions are available. Hindi and Telugu audio files have sampling frequencies of 8 and 16 kHz, respectively. Both have 16-bit encoding. Telugu audio is downsampled to 8 kHz in our experiments. We randomly select 15,000 utterances (\(\approx\) 15 hours) from their train sets for training the domain-independent acoustic models. The senone labels required for training the acoustic model are obtained from the alignments generated by a HMM-GMM system trained using Kaldi. We also use a random selection of 1000 utterances from the test set to validate the domain independence of learned features. We refer to this set as *dev*. The Sanskrit dataset used in the target domain has 3395 utterances with 16 kHz sampling frequency and 16-bit encoding. The data is randomly divided into two sets-train and test, with approximately 5.5 hours (2837 utterances) in the train set and 1 hour (558 utterances) in the test set. The train set is used for domain training, and the test set is used for inference. The data is downsampled to 8 kHz before its use in all our experiments. The text corpus for building the Sanskrit language models makes use of the *wiki* Sanskrit data dump and data from several Sanskrit websites. The extracted text is cleaned to remove unwanted characters and pre-processed to restrict the graphemes to the Devanagari Unicode symbols.
## Details of feature extraction
We use 40-dimensional filterbank features together with their delta and acceleration coefficients. Cepstral mean variance normalisation is performed at the utterance level. The features are spliced with a left and right context of 5 frames each. Thus the acoustic feature vector at the input of the DNN has dimensions of 1320 (40 x 3 x 11). Feature extraction is performed using Kaldi.
## Training of the GRL model
The feature extractor (\(G_f\)) has six hidden layers with 1024 nodes in each layer. The input to the \(G_f\) is a 1320-dimensional acoustic feature vector, and the output is a 1024-dimensional feature \(\mathbf{f}\). The feature vector \(\mathbf{f}\) is forwarded to both the senone classifier \(G_y\) and the domain classifier \(G_d\). The senone classifier has two hidden layers, each with 1024 nodes and an output layer with 3080 nodes (equal to the number of senones in the Hindi training data). Domain classifier has a hidden layer with 256 nodes and an output layer with two nodes corresponding to the source and target domains. All the hidden layers are followed by batch normalization and ReLU activation. The logarithm of the softmax is computed at the output of both domain and senone classifiers. All the parameters (\(\theta_f\), \(\theta_y\) and \(\theta_d\)) are updated during the training with Hindi utterances. Only \(\theta_f\) and \(\theta_d\) are updated during the training with unlabelled Sanskrit utterances. Negative log-likelihood loss is used for training. The models are trained using stochastic gradient descent with momentum. We use a batch size of 32 and an initial learning rate of 0.01. The learning rate is scaled by a factor of 0.95 after every 20000 steps. Training is performed for 20 epochs. The same number of frames from the source and target domains are used for training at every epoch. The domain adaptation factor \(\alpha\) is gradually changed from 0 to 1 using the approach in.
## Training of the DSN model
Acoustic frames from the source and target domains, both of dimension 1320, are Input to the DSN. Private encoders for the source and target have four hidden layers with 512 nodes in each layer. The shared encoder has six hidden layers with 1024 nodes each. The senone and domain classifiers have the same architecture as the GRL model. All the hidden layers are followed by batch normalization and ReLU activation. The shared decoder has three hidden layers and an output layer with 1320 nodes. The hyperparameters \(\beta\), \(\gamma\), and \(\delta\) are chosen as 0.25, 0.075, and 0.1, respectively. In order to promote the learning of the senone classifier in the initial phase of training, domain adversarial similarity losses are activated only after 10000 steps. The rest of the training process is the same as in the GRL model.
## Decoding
During the inference stage, only the output of the senone classifier is considered. The pre-softmax output of the senone classifier is normalized using the log probability of priors. To find the most probable word sequence, we use weighted finite-state transducers (WFST) based decoding. The vocabulary of Sanskrit is distinct from that of Hindi. So the FSTs for grammar (G) and lexicon (L) are built using the text corpus collected in Sanskrit (target domain). Pronunciation dictionary for building the L-FST uses the grapheme to phoneme (G2P) mapping scheme in Sanskrit. This is different from the Hindi G2P scheme in aspects like schwa deletion and pronunciation of visargas. HMM (H) and context-dependency (C) FSTs are created using the HMMs learned from the source domain data. These four FSTs are composed to form a single HCLG graph, which maps the senones directly to the words in the target domain.
# Results {#sec:results}
We decode the Sanskrit test set of 558 utterances with the adversarially trained GRL and DSN architectures. These models use both the labeled Hindi data and the unlabelled Sanskrit data for training. In order to benchmark the performance of these UDA models, we decode the utterances using a simple DNN model trained only with the Hindi speech data. In this model, the domain classifier is not part of the network architecture. We also compare our results with a DNN model trained in multi-task (MT) learning setup, training the whole network to minimize both the senone and domain classification objectives. This network has the same architecture as the network in Fig. [\[fig:blockdiagram\]](#fig:blockdiagram){reference-type="ref" reference="fig:blockdiagram"}, except that it does not have the GRL. This model also makes use of both the Hindi and Sanskrit data.
### Effect of the loss functions in DSN
Next, we experiment with the constituents of the loss function in DSN. We train four models: (i) with all the loss functions in DSN, (ii) without the difference loss (\(\gamma\) = 0), (iii) without the similarity loss (\(\beta\) = 0), and (iv) with the reconstruction loss computed as SIMSE. The results are listed in Table [7](#tab:dsnloss){reference-type="ref" reference="tab:dsnloss"}. The performance degrades slightly in the absence of difference loss which tries to enhance the orthogonality between the private and shared components. There is considerable degradation in the performance in the absence of similarity loss. The results are still better than the baseline DNN and MT models (refer column I in Table [5](#tab:wer){reference-type="ref" reference="tab:wer"}) indicating the usefulness of private and shared component decomposition in DSN. The model using SIMSE as the reconstruction loss performs inferior to that of the one using MSE.
[\[tab:telugu\]]{#tab:telugu label="tab:telugu"}
# Conclusions {#sec:conclusions}
In this work, we propose UDA as an option to tackle the scarcity of data in low-resource languages which share a common acoustic space with a high-resource language. We experiment with Hindi as the source domain language and Sanskrit as the target domain language. GRL and DSN models improve the WER by 6.71% and 7.32%, respectively, compared to a baseline DNN model trained only on Hindi. The models perform better than the multi-task learning framework. Proper selection of source domain language (Telugu in our case) further improves the results. The results indicate that UDA can provide a faster way of building ASR systems in low-resource languages, reducing the hassle of collecting large amounts of annotated training data if a suitable high-resource language is available.
# Introduction {#sec:intro}
Advancements in deep learning have brought performance improvements in acoustic and language modeling, yielding robust automatic speech recognition (ASR) systems in many languages. However, such systems require a large amount of speech data and the associated transcriptions. It is tough to collect large volumes of paired speech and transcriptions for most low-resource languages. It is estimated that only about 1% of the world languages have the minimum amount of data that is needed to train an ASR. However, in many cases, especially for languages in south Asia, we can find a "close enough" language with the same set (or a superset) of phonemes as the low-resource language and enough resources for building an ASR. In this work, we show that a better performing ASR can be built for the low-resource language using unsupervised domain adaptation (UDA) of acoustic models from the corresponding high-resource language. This method has the benefit of modeling on real data in comparison to the data augmentation techniques like vocal tract length perturbation, speech and tempo perturbation, noise addition, data synthesis, and spectral augmentation, where the modeling makes use of the synthetic data as well. Unsupervised domain adaptation (UDA) has been successfully applied to various tasks to alleviate the shift between the train and test distributions. shows good adaptation performance in the classification task on digit image datasets having considerable domain shifts. They learn features that are discriminative for the image classification task and invariant to the domain. They introduce a gradient reversal layer (GRL) for achieving this objective. reports performance improvements in speech recognition for data shifted in the domain by gender and accent. employs GRL layers to reduce the mismatch between train and test domains in the task of emotion recognition from speech data. uses the GRL approach to improve the word error rate (WER) in speech recognition with the source domain data as clean speech and the target domain data as contaminated speech. They also show the robustness of the approach to the domain shifts caused by the differences in datasets. The basic UDA scheme with a GRL tries to learn domain invariant features but ignores the individual characteristics of each domain. introduces domain separation networks (DSN) and shows improvements in a range of UDA scenarios in the image classification task. They learn two representations: one specific to each domain and the other common to both domains. uses DSN for adaptation from clean speech to noisy speech. In this work, we explore the feasibility of the UDA schemes in the ASR task on low-resource languages that share the same acoustic space with a high-resource language. Specifically, we place the following assumptions in the selection of high and low resource language pairs.
1. The acoustic space spanned by the low-resource language is a subspace of that of the high-resource language. This requires the phoneme set of the low-resource language to be a subset of the high-resource language.
2. There exists a high-resource language with reasonable amount of paired audio data and transcriptions to build an ASR.
3. The low-resource language has enough text data available to train the language models.
4. The speech data available in the low-resource language is quite limited, and the transcriptions are not available.
Though the above assumptions seem quite constrictive, we can easily find a few low-resource languages in the Indian subcontinent which share a common acoustic space with a reasonably high-resource language. In this work, we use Hindi as the high-resource language and Sanskrit as the low-resource language, both belonging to the Indo-Aryan language family. Hindi is written in the Devanagari script, and many of its words are derived from Sanskrit. However, there exist substantial differences in the vocabulary and pronunciation between the two languages. One of the important differences is that the schwa, implicit in each consonant of the script, is not pronounced at the end of words and in some other contexts in Hindi. The script does not tell us when the schwa should be deleted, a phenomenon known as schwa deletion. For example, the word for \"salty\" is pronounced as *nam'kīn* in Hindi and not *namakīna*. Another difference is the pitch accents that are common in Sanskrit. Despite the above differences, these languages share a common phoneme set. So we can intuitively argue that there exists a domain shift between the distributions of acoustic features in Hindi and Sanskrit. This makes us believe that the speech recognition problem in Sanskrit, an extremely low-resource language, may be posed as an UDA problem from Hindi, a language with a reasonably good collection of annotated audio.
# Unsupervised domain adaptation for acoustic modeling {#sec:uda}
We pose the problem of acoustic modeling in Sanskrit as an unsupervised domain adaptation task from Hindi. We build a deep neural network (DNN)-hidden Markov model (HMM) ASR system for Sanskrit with the domain-independent acoustic models learned from Hindi through UDA approaches. We make use of the UDA schemes introduced in and.
## Adversarial training using GRL {#sec:grl}
In GRL-based adversarial training, we try to learn a feature representation invariant to the domain, but good enough in discriminating the senone labels. The neural network architecture for learning the acoustic model consists of three parts: feature extractor, senone classifier, and domain classifier. A block diagram of the UDA architecture employing GRL is shown in Fig. [\[fig:blockdiagram\]](#fig:blockdiagram){reference-type="ref" reference="fig:blockdiagram"}. Feature extractor \(G_f\) maps the input acoustic features \(\mathbf{x}\) to an internal representation \(\mathbf{f} \in \mathcal{R}^D\). Senone classifier \(G_y\) maps the output of the feature extractor to the senone labels \(\mathbf{y}\) whereas the domain classifier \(G_d\) maps them to domain labels \(\mathbf{d}\). We train the network to minimize the senone classification loss during the training phase by optimizing the parameters in the feature extractor \(G_f\) and senone classifier \(G_y\). This makes the network look for features that are capable of discriminating the senone labels. To make the features domain-invariant, we optimize the parameters of the feature extractor to maximize the domain classification loss. But at the same time, the parameters of the domain classifier \(G_d\) are optimized to minimize the domain classification loss. This is achieved by introducing a GRL between the feature extractor \(G_f\) and domain classifier \(G_d\). During the forward pass, GRL acts as an identity transform. GRL takes the gradient from the subsequent layer during the backward pass, multiplies it with \(-\alpha\), a hyperparameter to control the trade-off between senone discrimination and domain-invariance, and passes it to the preceding layer. During the inference time, the domain classifier and GRL are ignored. The acoustic feature vectors are passed through the feature extractor and the senone classifier, and senone labels are predicted.
## Domain separation networks {#sec:dsn}
A block diagram of the DSN architecture is shown in Fig. [\[fig:dsn\]](#fig:dsn){reference-type="ref" reference="fig:dsn"}. They model both the private and shared components of the domain representation. Private encoders \(E_p^t(.)\) and \(E_p^s(.)\) extract components \(f_p^t\) and \(f_p^s\) which are specific to the target and source domains. Shared encoder \(E_c(.)\) is common to both domains and extract the shared components \(f_c^t\) and \(f_c^s\). Shared decoder \(D\) tries to reconstruct the input using the private and shared components. \(G(.)\), the senone classifier maps \(f_c^s\) to the senone label \(\mathbf{\hat{y}}\). The domain classifier \(Z(.)\) maps the shared components \(f_c^s\) and \(f_c^t\) to their respective domain labels \(\mathbf{\hat{d}}\). The network is trained to minimize the following loss function with respect to the parameters of \(E_p\), \(E_c\), \(G\) and \(Z\): \[L = L_{class} + {\beta} L_{sim} + {\gamma} L_{diff} + {\delta} L_{recon}\] where \(\beta, \gamma\), and \(\delta\) are hyperparameters. \(L_{class}\) represents the senone classification loss and is applied only to the source domain. It is computed as the negative log-likelihood of the ground-truth senone labels. \(L_{sim}\) represents the domain adversarial similarity loss and is computed as the negative log-likelihood of the domain labels. \(L_{sim}\) ensures that the shared components \(f_c^t\) and \(f_c^s\) are as similar as possible irrespective of their domain so that the domain classifier cannot reliably predict the domain of the sample from its shared representation. Parameters of the domain classifier \(Z(.)\) are trained to minimize the domain classification loss while the parameters of the shared encoder \(E_c\) are trained to maximise the domain classification loss. This is also accomplished with a GRL. \(L_{diff}\) encourages the shared component \(f_c\) and the private component \(f_p\) to encode different aspects of the input. This is achieved by imposing a soft subspace orthogonality constraint between the private and shared components. \[L_{diff} = \| \mathbf{F}_c^{sT} \mathbf{F}_p^s \|^2_F + \| \mathbf{F}_c^{tT} \mathbf{F}_p^t \|^2_F\] where \(\mathbf{F}_c^{s}\), \(\mathbf{F}_c^{t}\), \(\mathbf{F}_p^s\) and \(\mathbf{F}_p^t\) are matrices with \(f_c^s\), \(f_c^t\), \(f_p^s\) and \(f_p^t\) as rows. \(\|.\|_F\) denotes the Frobenius norm. \(L_{recon}\) is the reconstruction loss, computed as the mean squared error (MSE) between \(\mathbf{x}\) and \(\hat{\mathbf{x}} = D (E_c(\mathbf{x})+E_p(\mathbf{x}))\). \[L_{recon} = \sum_{i=1}^{N_s} \| \mathbf{x}_i^s-\hat{\mathbf{x}}_i^s\|^2 + \sum_{i=1}^{N_t} \| \mathbf{x}_i^t-\hat{\mathbf{x}}_i^t\|^2\] where \(N_s\) and \(N_t\) represent the number of speech frames from the source and target domains. We also validate the performance of the system with scale-invariant mean squared error (SIMSE) which is computed as: \[L_{SIMSE} = \frac{1}{k} \| \mathbf{x}-\hat{\mathbf{x}} \|_2^2-\frac{1}{k^2} (\left[\mathbf{x}-\hat{\mathbf{x}}\right]. \mathbf{1}_k)^2\] where \(k\) in the dimension of the input vector \(\mathbf{x}\), \(\mathbf{1}_k\) is a \(k\)-dimensional vector of ones and \(\|.\|_2\) is the \(L_2\) norm.
# Experimental setup {#sec:setup}
## Datasets used for the study
We primarily use a Hindi dataset in the source domain. We also use a Telugu dataset in the ablation studies. Both these datasets are the same as the ones used in multilingual and code-switching ASR challenge, Interspeech-2021. The details of both these datasets are available in. Both the speech data and the corresponding transcriptions are available. Hindi and Telugu audio files have sampling frequencies of 8 and 16 kHz, respectively. Both have 16-bit encoding. Telugu audio is downsampled to 8 kHz in our experiments. We randomly select 15,000 utterances (\(\approx\) 15 hours) from their train sets for training the domain-independent acoustic models. The senone labels required for training the acoustic model are obtained from the alignments generated by a HMM-GMM system trained using Kaldi. We also use a random selection of 1000 utterances from the test set to validate the domain independence of learned features. We refer to this set as *dev*. The Sanskrit dataset used in the target domain has 3395 utterances with 16 kHz sampling frequency and 16-bit encoding. The data is randomly divided into two sets-train and test, with approximately 5.5 hours (2837 utterances) in the train set and 1 hour (558 utterances) in the test set. The train set is used for domain training, and the test set is used for inference. The data is downsampled to 8 kHz before its use in all our experiments. The text corpus for building the Sanskrit language models makes use of the *wiki* Sanskrit data dump and data from several Sanskrit websites. The extracted text is cleaned to remove unwanted characters and pre-processed to restrict the graphemes to the Devanagari Unicode symbols.
## Details of feature extraction
We use 40-dimensional filterbank features together with their delta and acceleration coefficients. Cepstral mean variance normalisation is performed at the utterance level. The features are spliced with a left and right context of 5 frames each. Thus the acoustic feature vector at the input of the DNN has dimensions of 1320 (40 x 3 x 11). Feature extraction is performed using Kaldi.
## Training of the GRL model
The feature extractor (\(G_f\)) has six hidden layers with 1024 nodes in each layer. The input to the \(G_f\) is a 1320-dimensional acoustic feature vector, and the output is a 1024-dimensional feature \(\mathbf{f}\). The feature vector \(\mathbf{f}\) is forwarded to both the senone classifier \(G_y\) and the domain classifier \(G_d\). The senone classifier has two hidden layers, each with 1024 nodes and an output layer with 3080 nodes (equal to the number of senones in the Hindi training data). Domain classifier has a hidden layer with 256 nodes and an output layer with two nodes corresponding to the source and target domains. All the hidden layers are followed by batch normalization and ReLU activation. The logarithm of the softmax is computed at the output of both domain and senone classifiers. All the parameters (\(\theta_f\), \(\theta_y\) and \(\theta_d\)) are updated during the training with Hindi utterances. Only \(\theta_f\) and \(\theta_d\) are updated during the training with unlabelled Sanskrit utterances. Negative log-likelihood loss is used for training. The models are trained using stochastic gradient descent with momentum. We use a batch size of 32 and an initial learning rate of 0.01. The learning rate is scaled by a factor of 0.95 after every 20000 steps. Training is performed for 20 epochs. The same number of frames from the source and target domains are used for training at every epoch. The domain adaptation factor \(\alpha\) is gradually changed from 0 to 1 using the approach in.
## Training of the DSN model
Acoustic frames from the source and target domains, both of dimension 1320, are Input to the DSN. Private encoders for the source and target have four hidden layers with 512 nodes in each layer. The shared encoder has six hidden layers with 1024 nodes each. The senone and domain classifiers have the same architecture as the GRL model. All the hidden layers are followed by batch normalization and ReLU activation. The shared decoder has three hidden layers and an output layer with 1320 nodes. The hyperparameters \(\beta\), \(\gamma\), and \(\delta\) are chosen as 0.25, 0.075, and 0.1, respectively. In order to promote the learning of the senone classifier in the initial phase of training, domain adversarial similarity losses are activated only after 10000 steps. The rest of the training process is the same as in the GRL model.
## Decoding
During the inference stage, only the output of the senone classifier is considered. The pre-softmax output of the senone classifier is normalized using the log probability of priors. To find the most probable word sequence, we use weighted finite-state transducers (WFST) based decoding. The vocabulary of Sanskrit is distinct from that of Hindi. So the FSTs for grammar (G) and lexicon (L) are built using the text corpus collected in Sanskrit (target domain). Pronunciation dictionary for building the L-FST uses the grapheme to phoneme (G2P) mapping scheme in Sanskrit. This is different from the Hindi G2P scheme in aspects like schwa deletion and pronunciation of visargas. HMM (H) and context-dependency (C) FSTs are created using the HMMs learned from the source domain data. These four FSTs are composed to form a single HCLG graph, which maps the senones directly to the words in the target domain.
# Results {#sec:results}
We decode the Sanskrit test set of 558 utterances with the adversarially trained GRL and DSN architectures. These models use both the labeled Hindi data and the unlabelled Sanskrit data for training. In order to benchmark the performance of these UDA models, we decode the utterances using a simple DNN model trained only with the Hindi speech data. In this model, the domain classifier is not part of the network architecture. We also compare our results with a DNN model trained in multi-task (MT) learning setup, training the whole network to minimize both the senone and domain classification objectives. This network has the same architecture as the network in Fig. [\[fig:blockdiagram\]](#fig:blockdiagram){reference-type="ref" reference="fig:blockdiagram"}, except that it does not have the GRL. This model also makes use of both the Hindi and Sanskrit data.
### Effect of the loss functions in DSN
Next, we experiment with the constituents of the loss function in DSN. We train four models: (i) with all the loss functions in DSN, (ii) without the difference loss (\(\gamma\) = 0), (iii) without the similarity loss (\(\beta\) = 0), and (iv) with the reconstruction loss computed as SIMSE. The results are listed in Table [7](#tab:dsnloss){reference-type="ref" reference="tab:dsnloss"}. The performance degrades slightly in the absence of difference loss which tries to enhance the orthogonality between the private and shared components. There is considerable degradation in the performance in the absence of similarity loss. The results are still better than the baseline DNN and MT models (refer column I in Table [5](#tab:wer){reference-type="ref" reference="tab:wer"}) indicating the usefulness of private and shared component decomposition in DSN. The model using SIMSE as the reconstruction loss performs inferior to that of the one using MSE.
[\[tab:telugu\]]{#tab:telugu label="tab:telugu"}
# Conclusions {#sec:conclusions}
In this work, we propose UDA as an option to tackle the scarcity of data in low-resource languages which share a common acoustic space with a high-resource language. We experiment with Hindi as the source domain language and Sanskrit as the target domain language. GRL and DSN models improve the WER by 6.71% and 7.32%, respectively, compared to a baseline DNN model trained only on Hindi. The models perform better than the multi-task learning framework. Proper selection of source domain language (Telugu in our case) further improves the results. The results indicate that UDA can provide a faster way of building ASR systems in low-resource languages, reducing the hassle of collecting large amounts of annotated training data if a suitable high-resource language is available. | {'timestamp': '2021-09-17T02:12:38', 'yymm': '2109', 'arxiv_id': '2109.05494', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05494'} |
null | null |
# Introduction {#sec:intro}
Quantum hardware available today with fifty-plus qubits can already outperform the world's most advanced supercomputer for certain problems . In the near-term, we can expect quantum computers with few hundreds of qubits to solve certain domain-specific applications . Unfortunately, the fidelity of applications executed on these *Noisy Intermediate Scale Quantum (NISQ)* computers is limited by the high error-rates of the physical qubit devices. The probability of encountering an error on NISQ computers increases with the size of the program. Therefore, developing software solutions that can reduce the impact of hardware errors and improve the fidelity of NISQ applications is an active area of research. A qubit can encounter errors while performing gate or measurement operations. Additionally, a qubit can also accumulate errors while it is idle and not performing any operations. These errors, referred to as *idling errors*, are observed on both superconducting and ion-trap hardware . Under certain circumstances, the idling error-rate of a qubit can exceed the error-rate from gate operations. Furthermore, idling errors can increase significantly in the presence of other active qubits in the vicinity, as the idle qubit accumulates phase noise due to crosstalk generated by the on-going gate operations. Our experiments on IBMQ hardware show that an idle qubit is almost 10x more vulnerable to errors when two-qubit gate operations are scheduled adjacent to it. Thus, idling errors can significantly degrade the fidelity of quantum programs and we focus on mitigating these errors in this paper. The characteristics of quantum programs and NISQ hardware cause a large number of program qubits to remain idle during the execution. There are three key reasons for qubits to remain idle: (1) limited parallelism, (2) high latency of two-qubit gates, and (3) additional data movement due to SWAP operations. Quantum programs have limited operational parallelism as complex multi-qubit operations are translated into a highly serial sequence of two-qubit CNOT gates. Similarly, there exists significant non-uniformity in the latency of different operations on NISQ computers. For example, the latency of a CNOT gate on IBMQ hardware (\(\approx\) 400 ns) is almost an order of magnitude higher than the latency of a single qubit gate. Also, CNOT gates on the same hardware incur different latencies. For example, while the latency of a CNOT gate is 440 ns on average, it can be as high as 860 ns on IBMQ-Toronto. Therefore, even if a program can orchestrate parallel operations on different qubits, some with single-qubit operations and others with two-qubit operations, the qubits with single-qubit operations finish execution earlier than the qubits with two-qubit operations and remain idle. Similarly, parallel two-qubit gates with variable latencies finish execution at different times. Finally, architectural constraints can also cause idle times in programs as current machines do not have all-to-all connectivity. When two unconnected qubits need to perform a two-qubit operation, the compiler inserts SWAP instructions (typically performed as a sequence of 3 CNOT instructions) to perform data movement, which causes serialization and large idle periods due to the long latency of these operations. By keeping the idle qubits busy with a specific sequence of gates, their susceptibility to spurious noise can be reduced. *Dynamical Decoupling (DD)* is a well-known technique from the quantum devices community that uses this insight to mitigate idling errors. DD is implemented by the repeated execution of a sequence of single-qubit operations that return the qubit to its original state. Thus, DD operations do not change the overall state of the qubits as they collectively behave as an identity gate that suppresses noise. DD is widely used in device-level characterization experiments to remove systematic noise . More recently, DD has been adopted in quantum volume and quantum error correction circuits . So far, the use of DD has mainly been limited to single qubit experiments or specific circuits . Most importantly, at the application-level, prior studies have relied on enabling DD for all idle qubits. However, our experiments on IBMQ systems show that such indiscriminate application of DD often results in sub-optimal improvements in fidelity because while DD can reduce idling errors, it can still increase the effective error-rate of a qubit if the errors introduced due to the extra operations outweighs the benefits. Moreover, the effectiveness of DD depends on the structure of the program, qubit state, and device characteristics. We explain this dependency with an example. Figure [\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(a) shows a quantum program with 3-qubits. Note the difference in instruction latencies (40ns for \(H\) versus 400ns and 600ns for the \(CNOT\)s) and that qubits \(\mathsf{q}[0]\) and \(\mathsf{q}[2]\) remain idle for significant periods of time. Figure [\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(b-d) shows the three options for applying DD--to all idle qubits, only \(\mathsf{q}[0]\), and only \(\mathsf{q}[2]\), respectively. Figure [\[fig:intro\]](#fig:intro){reference-type="ref" reference="fig:intro"}(e) compares the reliability of the baseline with the three options for DD. While applying DD for all idle qubits provides some reliability benefits, the highest improvement comes when DD is applied to only a subset of the qubits (only \(\mathsf{q}[2]\) in this example). Thus, to mitigate idling errors at the application-level, DD must be applied robustly and judiciously for the most optimal performance. To that end, this paper proposes *Adaptive Dynamical Decoupling (ADAPT)*, a software framework to reliably use DD for NISQ applications by identifying the subset of qubits that are most likely to benefit from DD. ADAPT employs a trial-and-error method to search for the subset of qubits that maximizes the application fidelity. However, the output of the program must be known a-priori for this search to be effective, which is not possible. We observe that the effectiveness of DD depends on the program structure and two programs with similar structures tend to have similar fidelity when executed on identical physical qubits. ADAPT leverages this insight and searches for the optimal DD sequence using a *Decoy Circuit* that is structurally similar to the given program but built using Clifford operations (and only a handful of non-Clifford instructions). Clifford gates (CNOT, H, X, Z, S) can be efficiently simulated on conventional computers and therefore, the noise-free output of the decoy circuit can be estimated. The decoy circuit shows a similar trend in idling errors as the input circuit because it uses identical CNOT operations as the input circuit. ADAPT performs the trial-and-error search on the decoy circuit by inserting different DD sequences on the subset of the qubits and selecting the pattern that maximizes the likelihood of getting the correct answer on the decoy circuit. While the number of possible DD combinations increases exponentially with the number of qubits, ADAPT employs a divide-and-conquer approach where only 4 qubits are evaluated exhaustively at any time, keeping the search tractable. This results in a linear complexity in the number of qubits. We evaluate the effectiveness of ADAPT using IBMQ systems, ranging from 16 to 27 qubits, and two different DD protocols. Our evaluations show that ADAPT is robust and improves fidelity for both types of DD sequences, making it generalizable to other systems and DD protocols. ADAPT improves the fidelity of key NISQ benchmarks on average by 1.86x and by up-to 5.73x compared to the baseline without DD, and by 1.2x compared to applying DD to all qubits. The software for ADAPT and datasets for the evaluations in this paper is available at this [[link]{.ul}](https://github.com/pdas36/ADAPT). Overall, this paper makes the following contributions:
1. To the best of our knowledge, this is the first paper to evaluate DD at an application level. We show that while DD is beneficial in general, applying DD indiscriminately to all the idle qubits does not provide the highest fidelity.
2. We propose ADAPT, a software framework that applies DD judiciously by estimating the subset of qubits that are likely to provide the highest reliability with DD.
3. We present a Clifford-based decoy circuit approach and an efficient search algorithm to practically implement ADAPT.
# Background
## Quantum Bits and Gates
The state of a qubit is denoted as a superposition of its basis states \(\ket{0}\) and \(\ket{1}\), i.e., \(\ket{\psi} = \alpha\ket{0} + \beta\ket{1}\). It can be represented as a point on the Bloch sphere, as shown in the Figure [\[fig:back\]](#fig:back){reference-type="ref" reference="fig:back"}(a). When a qubit with state \(\ket{\psi}\) is measured, it produces \"0\" with a probability of \(\alpha^2\) and \"1\" with probability \(\beta^2\). A single qubit gate rotates the qubit state from one point on the sphere to another, as shown in Figure [\[fig:back\]](#fig:back){reference-type="ref" reference="fig:back"}(a).
## NISQ Hardware Errors {#sec:nisqerrors}
For a qubit in superposition, even a tiny change in its energy can produce a valid but *erroneous* state. The likelihood of undesirable changes in the state of a qubit is defined as the qubit error rate. Qubit errors can broadly be classified into: These errors are caused when the qubit is performing gate or measurement operations. On current generation of quantum computers from IBM, the error rate is approximately 0.1% for single-qubit operations, 1%-2% for two-qubit operations, and 4% for measurement operations. These errors occur when a qubit is idle and not performing any operations. For example, coherence errors can cause qubits to naturally decay to the lowest energy state (\(\ket{0}\)) within a short time (\(10-100\) \(\mu\textrm{seconds}\)). Moreover, qubits can lose their phase information due to their interactions with the environment (dephasing). Similarly, undesirable crosstalk can cause operations on other active qubits to affect the state of neighboring (spectator) idle qubits. Idling errors can also be caused by environmental noise.
## NISQ Model for Quantum Computing
Hardware errors can cause a quantum program to produce an incorrect answer. Unfortunately, in the near term, quantum computers will not have enough resources to perform error correction (which can incur 20x-100x overhead). However, some quantum algorithms can tolerate limited hardware errors at the algorithmic level and can be used to solve practical problems using the *Noisy Intermediate-Scale Quantum (NISQ)* model, wherein the program is run thousands of times to identify the correct answer. NISQ systems can solve problems that are beyond the reach of existing computers . However, the ability to infer the correct answer on NISQ machines depend on the error-rates and program length. Recent works have proposed software solutions to reduce the length of programs and use error characteristics to improve program fidelity . Although useful in mitigating measurement, gate errors, and CNOT-CNOT crosstalk errors , these schemes do not focus on idling errors. In this paper, we focus on software policies to mitigate idling errors at application-level.
## Idling Errors in NISQ Applications
The characteristics of quantum programs cause many of the qubits to remain idle for a significant period of time during program execution. There are three key reasons for qubits to remain idle: (1) limited parallelism, as quantum programs get decomposed into sequence of instructions with data dependency (2) high latency of two-qubit gates compared to single-qubit gates, and (3) additional data movement or SWAP operations. For example, let us take a 4-qubit Bernstein-Vazirani (BV) circuit shown in Figure [\[fig:bv_variable\]](#fig:bv_variable){reference-type="ref" reference="fig:bv_variable"}(a). Due to low parallelism, the CNOT gates of this circuit must be scheduled serially. Consequently, qubit Q0 remains idle when CNOTs and are executed. Although existing compilers minimize idle times by scheduling instructions *as late as possible* , this optimization is not feasible for all qubits as the computation must make forward progress. For example, late initialization causes Q2 to not experience any idle time but cannot optimize Q1 from remaining idle during the execution of CNOT . The long latency of the CNOT gates exacerbates the idle times. Even if a program can orchestrate parallel operations on different qubits, qubits with single-qubit operations (10x faster) finish execution earlier than qubits with two-qubit gate operations and remain idle. Furthermore, CNOT gates on the same hardware incur different latencies. For example, the worst-case CNOT gate latency on IBMQ-Toronto is 1.95x the average latency. Consequently, parallel CNOT gates with variable latencies finish at different times. Finally, NISQ compilers insert SWAP instructions to overcome limited device connectivity causing serialization and long idle periods. For example, Figure [\[fig:bv_variable\]](#fig:bv_variable){reference-type="ref" reference="fig:bv_variable"}(b) shows that the idle time of qubit Q0 for different BV circuits is about an order of magnitude higher on IBMQ-Toronto compared to a machine with similar error rates but all-to-all connectivity.
[\[tab:idle\]]{#tab:idle label="tab:idle"}
Idling errors depend on the state of qubits and concurrent CNOTs. Further, the effectiveness of DD change across calibration cycles. For example, Figure [\[fig:downsideofdd\]](#fig:downsideofdd){reference-type="ref" reference="fig:downsideofdd"} shows the relative fidelity of Qubit-12 when CNOT operations are performed on the Link:17-18 for two different calibration cycles. In the first cycle, DD improves the fidelity up-to 1.27x, whereas DD degrades the fidelity up-to 0.35x in the second cycle. Our experiments also show that idling errors exist between qubit-link pairs that may not be present in the same on-chip neighborhood, making localized characterization approaches inadequate. We make similar observations using other (a) systems such as IBMQ-Paris and IBMQ-Casablanca and (b) DD pulses .
The characterization complexity increases further when simultaneous CNOT operations are considered on multiple links. Our experiments show that although CNOT operations on multiple links can result in higher idling errors in general, such additive effect does not exist always. For some of the cases, one of the active links dictates the overall idling error. In rare situations, multiple CNOTs can reduce idling errors. To summarize, our evaluations using IBMQ systems show complex trends in idling errors and effectiveness of DD. We confirm that (1) on-going CNOTs increase idling errors, (2) the idling error-rate and effectiveness of DD depends on the CNOT patterns and combination of the idle and active qubits, and (3) the idle duration.
## Impact of DD on Application Fidelity
The most straightforward method to enable dynamical decoupling at the application level is to insert DD pulses wherever feasible. To implement this design, we can identify all program regions where each qubit is idle and insert DD sequences. However, naively inserting DD sequences for all qubits may not always be beneficial, as we have already observed from the characterization experiments. To demonstrate this effect at the application-level, we execute two 6-qubit benchmarks, *Quantum Fourier Transform (QFT)* and *Bernstein Vazirani (BV)*, on 27-qubit IBMQ-Toronto, with DD applied on all 64 (\(2^6\)) possible qubit combinations. Figure [\[fig:qft_adder_all_dd\]](#fig:qft_adder_all_dd){reference-type="ref" reference="fig:qft_adder_all_dd"} shows the fidelity (likelihood of getting correct answer) for all 64 DD combinations, with 0 (000000) being the baseline when no DD is applied, and 63 (111111) being the case when DD is applied on all six qubits. We observe that both benchmarks show significant variation in fidelity for different DD sequences. For *QFT*, enabling DD for all qubits increases the fidelity by 2.6x, however, the fidelity can be improved up-to 6.6x by choosing sequence \"010100\". For *BV*, applying DD on all qubits degrades fidelity to 0.88x, and we may deem DD to be counter-effective for this benchmark. However, using the DD sequence \"010100\" can improve the fidelity by 1.1x compared to no-DD and up-to 1.26x compared to DD-for-all. Note that the best DD sequence depends on the workload characteristics and the physical qubits used to run the program.
# Adaptive Dynamical Decoupling
To enable the robust use of DD at the application level, we propose *Adaptive Dynamical Decoupling (ADAPT)*. ADAPT identifies the combination of DD sequences that suppress idling errors and maximizes the application fidelity. ADAPT is implemented as a compiler pass that can be easily integrated with existing and future quantum compiler tool flows. In this section, we provide an overview of ADAPT, discuss the design issues in estimating the optimal DD sequence, propose scalable search algorithms for the same.
## Overview of ADAPT
ADAPT identifies all idle qubit slots in a quantum circuit and applies DD gate sequences during these idle periods. However, the most optimal subset of qubits on which DD must be applied is neither known a-priori to program execution nor practically feasible to obtain through extensive device characterization. To overcome this challenge, ADAPT relies on a *Decoy Circuit* which is structurally similar to the input program, but with a known solution. Furthermore, to limit the complexity of the search for the optimal DD sequence, ADAPT employs a localized algorithm. Figure [\[fig:overview\]](#fig:overview){reference-type="ref" reference="fig:overview"} shows an overview of ADAPT which accepts a quantum circuit as the input and outputs the circuit with the most optimal DD sequence. We discuss the specific design details of ADAPT next.
## Clifford Decoy Circuits (CDC)
If the outcome of a quantum circuit is known, we can apply DD on different subsets of qubits and assess their effectiveness in improving fidelity. Unfortunately, practical applications are hard to simulate using conventional computers, and their correct outcome is unknown. To overcome this challenge in estimating the optimal DD sequence, we leverage the following insights. *Insight \#1: Not all quantum circuits are hard to simulate and circuits comprising of Clifford gates only can be simulated efficiently on conventional computers .* *Insight \#2: Our characterization experiments show that crosstalk from CNOT operations is a dominant source of idling errors. Thus, two circuits with similar CNOT structures encounter similar idling errors.* ADAPT uses these two insights to generate an efficient *Clifford Decoy Circuit (CDC)* that preserves the structure of the input program. Next, ADAPT applies different DD combinations to this decoy circuit and selects the DD sequence that maximizes the fidelity of the decoy circuit. Finally, ADAPT applies this optimal DD sequence to the input circuit and executes it.
### Design of Clifford Decoy Circuits (CDC)
\
[\[sec:cliffordreplacement\]]{#sec:cliffordreplacement label="sec:cliffordreplacement"} ADAPT relies on Clifford Decoy Circuits generated using gates from the Clifford group--*CNOT, X, Y, Z, H, S*. To create the CDC of a circuit, ADAPT replaces the non-Clifford gates of the circuit using the closest Clifford gates. To measure the closeness of a non-Clifford gate in the program with a Clifford gate, we use an operator norm-a distance measure used in the literature for approximating one unitary with another. \[\begin{aligned}
\|U-V\|_{\infty}:=\max _{|\psi\rangle \neq 0} \frac{\|(U-V)|\psi\rangle \|_{2}}{\||\psi\rangle \|_{2}} \end{aligned}\] For example, by using operator norm, the U1 gate is either replaced by Z or S gates, whereas U2 and U3 gates are replaced by the closest Clifford gates depending on the Euler angles associated with the gates. As CNOT is a Clifford gate, the structure and usage of these gates are identical between the CDC and the input program and therefore, the CDC encounters similar crosstalk from CNOT operations. This also ensures that the qubits in the CDC experience similar idle times as the original circuit.
### Effectiveness of Clifford Decoy Circuits
\
To test the effectiveness of decoy circuits, we compare the fidelity of a 4-qubit quantum ADDER benchmark and its corresponding CDC for all possible DD sequence combinations. For example, this five qubit program has 16 (\(2^4\)) possible DD sequence combinations where the combination "0 (0000)\" indicates that DD is not applied on any of the qubits, whereas the combination "15 (1111)\" indicates that DD is applied to all of the qubits. We also compute the Spearman's Correlation Coefficient to quantify the agreement between the input program and the CDC. Figure [\[fig:adder_decoy_correlation\]](#fig:adder_decoy_correlation){reference-type="ref" reference="fig:adder_decoy_correlation"} shows the trend in program fidelity for the actual circuit and the CDC and we observe that the program fidelity is strongly correlated to the fidelity of the CDC (Spearman's Correlation Coefficient = 0.78)
### Overcoming the Limitations of CDCs: Seeded Decoy Circuits
\
ADAPT generates sub-optimal DD sequences when there is a mismatch between the fidelity trend of the input program and its CDC. Our experiments show that a CDC with high variance in the output distribution can be insensitive to changes in idling errors and relying on them may result in sub-optimal DD sequences. For example, if a CDC produces a uniform distribution, executing this CDC with different DD sequences does not significantly change the output distribution in the presence of idling errors. To tackle this problem, we propose the use of *Seeded Clifford Decoy Circuits (SDC)* that generate output distributions with low entropy, thus making them sensitive to idling errors. While simply removing all the single qubit gates from an input program and preserving the CNOT structure only, as shown in Figure [\[fig:sdc\]](#fig:sdc){reference-type="ref" reference="fig:sdc"}, can generate a decoy circuit, it does not truly mirror the fidelity trends of the input circuit because it does not capture the phase errors. To ensure that the output entropy is reduced while still being representative of the input circuit, SDCs use a very limited number of non-Clifford gates. SDCs apply an initial layer of non-Clifford gates on few qubits and replace the remaining non-Clifford gates with Clifford gates. For example, unlike the CDC shown in Figure [\[fig:sdc\]](#fig:sdc){reference-type="ref" reference="fig:sdc"}(c) that uses all Clifford gates, the SDC shown in Figure [\[fig:sdc\]](#fig:sdc){reference-type="ref" reference="fig:sdc"}(d) uses non-Clifford gates in the first layer of the circuit and Clifford gates in the later circuit layers. Our experiments show that SDCs can produce a rich state evolution while generating low entropy outputs. Although, the simulation cost of SDC is slightly higher than CDC, it is still not as expensive as running a full quantum simulation, which requires exponential resources. Moreover, SDCs produce low entropy outputs and thus, further optimizations to reduce the sampling cost can be deployed as well. For the purpose of simulations, we use Qiskit Extended Stabilizer Simulator (based on ). Table [2](#tab:SDC){reference-type="ref" reference="tab:SDC"} shows the effectiveness of SDCs in improving the correlation between the decoy and original circuits and the time required to simulate the SDCs for 64,000 shots. To test the scalability, we simulate a 100-qubit QAOA SDC, which requires 330 seconds for 100,000 shots. Note that CDCs or SDCs require simulation only once because applying different DD sequences does not alter the output on a simulator.
[\[tab:SDC\]]{#tab:SDC label="tab:SDC"}
[\[tab:ibmqplatforms\]]{#tab:ibmqplatforms label="tab:ibmqplatforms"}
[\[tab:benchmarks\]]{#tab:benchmarks label="tab:benchmarks"}
Prior works have used similar metrics such as Success Probability . We use distance-based metric as the output of the quantum program can be a probability distribution with multiple correct answers. Also, TVD closely matches prior metrics.
## Number of Trials
We perform experiments with up-to 32,000 shots, depending on the program size, to obtain the output probability distributions. The largest benchmark in our study uses ten qubits (QAOA-10), so it can produce a maximum of 1024 (\(2^{10}\)) unique solutions and therefore, the number of samples used in our study are sufficiently large for the workloads.
## Competing Policies
For our evaluations, we use four competing policies which are described next:
1. **No DD (Baseline)**: DD is not applied to any idle qubit.
2. **All-DD**: DD is applied on all the program qubits during any time period when they are idle.
3. **ADAPT**: The optimal DD sequence is obtained from a structured search using decoy circuits and applied.
4. **Runtime Best**: Evaluates the program with all possible DD sequences (\(2^N\) for an \(N\)-qubit program) and the sequence with highest fidelity at runtime is selected.
# Results
In this section, we provide the evaluation results for ADAPT across three different quantum computers: 27-qubit IBMQ-Paris, 27-qubit IBMQ-Toronto, and 16-qubit IBMQ-Guadalupe.
## Results for IBMQ-Paris
Figure [\[fig:Paris\]](#fig:Paris){reference-type="ref" reference="fig:Paris"} shows the Fidelity of four benchmarks for the XY4 protocol compared to the baseline (without DD). The number below each benchmark label specifies the baseline fidelity. We observe that on an average DD improves Fidelity. Applying DD on all qubits improves the fidelity by 1.97x on average and up to 2.89x. However, ADAPT improves the application fidelity by 3.27x and by up-to 5.7x. We also observe that the effectiveness of DD increases with increasing program size, which is expected because larger programs have more operations and depth, leaving room for longer idle time windows in the program. We also observe that the most optimal sequence at run-time outperforms both ADAPT and All-DD. This is due to the limitations of the decoy circuits and the limited search space explored by ADAPT. Our default ADAPT design searches the best sequence in a neighborhood of up-to four qubits at a time. For example, for QAOA-10, ADAPT uses only 36 decoy circuits, unlike the entire 1024 possible decoy circuits space. Nonetheless, we observe that the fidelity improvement of ADAPT is close to the runtime best and higher than All-DD for these workloads. We were unable to perform experiments using the IBMQ-DD protocol for IBMQ-Paris because of changes in the basis gates and subsequent retirement of the machine.
## Results for IBMQ-Toronto
Figure [\[fig:toronto_result\]](#fig:toronto_result){reference-type="ref" reference="fig:toronto_result"} shows the Fidelity of All-DD, ADAPT and the Runtime-Best policies for 27-qubit IBMQ-Toronto machine, relative to the baseline for two different DD protocols. The number below each benchmark label specifies the baseline fidelity of the application. The structure of the QFT circuit cause qubits remain idle for substantial time periods. For example, in `QFT-6B`, Qubit-0 is idle for 90% of the total time taken for the overall execution. Although a long sequence of DD gates adds a significant amount of single-qubit gate errors, it is still effective in improving the overall fidelity. Overall, ADAPT outperforms the baseline and improves the Fidelity by 1.52x and by up-to 3.1x for the XY4 protocol. Compared to All-DD, ADAPT improves the Fidelity on average by 1.3x and by up-to 1.89x. For the IBMQ-DD scheme, ADAPT improves the Fidelity by 1.47x and up-to 2.67x compared to the baseline. Thus, ADAPT is a generalized technique to identify qubits most vulnerable to idling errors at runtime and has applicability irrespective of the DD protocol.
## Results for IBMQ-Guadalupe
Figure [\[fig:guadalupe_result\]](#fig:guadalupe_result){reference-type="ref" reference="fig:guadalupe_result"} shows the Fidelity of the three different DD policies for the 16-qubit IBMQ-Guadalupe machine, normalized to the No-DD baseline. Note that this is one of the most recently released IBMQ systems with significantly reduced gate latencies and error-rates and improved coherence times. So, to test the robustness of ADAPT, we run slightly larger workloads (in terms of number of qubits, two-qubit operations, and circuit depth) on this machine. Here too, the number below each benchmark label specifies the baseline fidelity of the application. We observe that in general applying DD on all idle qubits for such large programs slightly degrades the fidelity in specific cases (`QFT-7A` for example). Note that the dominant source of errors in these circuits are not idling errors but gate and measurement errors. However, we observe that ADAPT is more robust and generally outperforms the All-DD policy. For example, the fidelity of the `QAOA-10B` benchmark improves by 3.1x compared to the baseline (without DD) and 4.65x compared to applying all-DD. However, we observe that in all these cases, the most optimal DD sequence at runtime outperforms the All-DD policy. Table [5](#tab:summary){reference-type="ref" reference="tab:summary"} summarizes the minimum, average, and maximum Fidelity for the three different dynamical policies normalized to the baseline on three IBMQ machines ranging from 16 to 27 qubits. Overall, ADAPT improves application Fidelity by 1.7x on average and up-to 5.73x compared to No-DD.
[\[tab:summary\]]{#tab:summary label="tab:summary"}
:::
## Impact of DD Pulse Type
We compare the effectiveness of the two protocols studied in this paper standalone using some additional characterization: the XY-4 sequence and the state-of-the-art XX sequence recently proven to be effective on IBMQ systems . In the experiment, we prepare three different circuits for variable idle times (T), as shown in Figure [\[fig:dd_sequence_comparison\]](#fig:dd_sequence_comparison){reference-type="ref" reference="fig:dd_sequence_comparison"}. In the circuit, a quantum state is prepared by performing a single qubit rotation and the associated qubit is kept idle. Throughout the idle time, CNOT operations are repeatedly performed on a specific physical link of the device that is not connected to the qubit under study. Finally, the qubit under study is brought back to the ground state by performing the inverse rotation. In the first circuit, no DD pulses are inserted, whereas in the second circuit the XY4 DD pulses inserted throughout the idle period. The third circuit uses IBM's DD sequence in which X(\(\pi\)) and X(\(-\pi\)) are evenly placed by waiting for specific delay slots, as shown in Figure [\[fig:dd_sequence_comparison\]](#fig:dd_sequence_comparison){reference-type="ref" reference="fig:dd_sequence_comparison"}(c). The time period for each individual delay slot is computed from the difference of idle time and length of the two X rotations, as described in Equation [\[eq:delayslot\]](#eq:delayslot){reference-type="eqref" reference="eq:delayslot"}. Note that we use the most optimal decomposition for both DD pulses to ensure a fair comparison. \[\label{eq:delayslot} \textrm{Delay }(\frac{\tau}{4}) = \frac{\textrm{T-length of X(}\pi\textrm{)-length of X(-}\pi\textrm{)}}{\textrm{4}}\] We run these circuits for each qubit and physical link combination (total of 224 combinations possible) on 16-qubit IBMQ-Guadalupe and Figure [\[fig:dd_sequence_comparison\]](#fig:dd_sequence_comparison){reference-type="ref" reference="fig:dd_sequence_comparison"}(d) shows the average fidelity of each circuit as the idle time is increased. We observe that on an average XY4 sequence outperforms the IBMQ DD sequence with increasing idle time. Similar results are reported for the Google Sycamore hardware as well as other works . This is because when idle periods are longer, there is still sufficient delay between the two X rotations during which errors can accumulate for the IBMQ-DD sequence. While such long idle periods may not be observed for random circuits used in Quantum Volume experiments, they exist in many other practical quantum applications (QFT for example). For circuits with long idling periods, we observe that the IBMQ DD protocol often performs worse than applying XY4 continuously because the latter does not experience large delays between DD pulses. To account for this in our evaluations at the application-level, we use the IBMQ-DD sequence in a more conservative manner by inserting the DD gate sequence multiple times for large idle periods. This enables a fair assessment of the DD protocol.
# Related Work
With the increasing number of qubits and improving device quality on near-term quantum architectures, quantum software can play a vital role in improving the reliability of NISQ applications . To improve the application fidelity, existing quantum compilers focus on minimizing the number of gates by searching for the best possible qubit mappings and sequence of SWAP operations. Moreover, recent works on noise-aware compilation use the underlying error characteristics of the hardware to avoid specific physical qubits and links such that the impact of worst-case errors on the application fidelity is reduced . Other works specifically target reducing certain types of errors such as crosstalk errors between ongoing CNOT operations, measurement errors . However, to the best of our knowledge, no compiler optimization so far focuses only on mitigating idling errors at the application level. As programs grow in size, the vulnerability of programs to idling errors increase and our paper focuses on tackling these errors. This is a well-studied noise mitigation technique that applies to a variety of qubit technologies . DD is ubiquitously used in qubit benchmarking and other small-scale experiments. More recently, Tripathi et al. have discussed the presence of crosstalk due to inter-qubit couplings and the role of DD in suppressing these errors on superconducting qubits . However, this study does not consider the crosstalk from CNOT operations which exists at the application level. Also, a generic theoretical framework for DD exists that can be used to integrate DD with fault-tolerant and noise mitigation protocols . While effective for calibration and well-studied from a theoretical perspective, the trade-offs in using DD at the application-level are not fully studied. Our work is inspired by the experimental demonstration of DD on IBM and Rigetti hardware in suppressing phase errors and improving coherence time (T2) using XY-4 and XX pulses . However, the use of DD to mitigate idling errors due to operational crosstalk, especially crosstalk generated by long latency CNOT gates, is not well explored. The recent study on IBMQ hardware only considers crosstalk from inter-qubit ZZ couplings . Recently, IBM demonstrated a milestone achievement of reaching a quantum volume of 64 and subsequently 128 for their superconducting machine, with DD as one of the components used in benchmarking the system. IBM used a custom DD protocol, wherein the compiler inserts DD pulses on all qubits wherever possible. This is similar to the *DD-on-All Qubits* configuration that we study as one of our baselines. We show that ADAPT is more effective than DD-on-All. DD on all qubits is also used in the study of quantum error correction codes (more specifically repetition codes) experiments on Google Sycamore device . Strikis et al. proposed an error mitigation strategy that inserts an extra gate before and after each operation to reduce both active and idle errors. They propose a learning scheme to identify the type of extra gates for error mitigation. Similarly, Zlokapa et al. propose to train a deep neural network to learn the noise characteristics of a 5-qubit machine and use this network to identify the best DD pulses. ADAPT is orthogonal to these existing approaches in that it tries to identify the subset of qubits that should use DD at runtime depending on the program and device characteristics. ADAPT can use the best DD sequence for that subset of qubits. While we use decoy circuits to find the best DD sequence in ADAPT, decoy circuits have been used in other scenarios as well, such as verifying cryptographic protocols .
# Conclusion
The quality and size of near-term quantum computers is improving. However, the device error-rates are still quite high and limit the fidelity of applications executed on them. In addition to facing errors while performing gate or measurement operations, qubits can also accumulate errors while remaining idle. Such idling errors present significant challenges in executing large programs. In this paper, we focus on mitigating idling errors for NISQ applications. Prior works have used *Dynamical Decoupling (DD)* to reduce idling errors by applying a sequence of DD gates when the qubit is idle. While DD has been shown to be effective at a small scale, its applicability at the application level is not yet fully studied. We show that applying DD to all the qubits in a program is sub-optimal and may even degrade the application fidelity in some specific cases because DD is implemented by introducing additional quantum gates. If the collective error-rate of these additional operations surpasses the idling error-rate, DD can adversely impact the overall fidelity at the application-level. Thus, to reduce the impact of idling errors at the application level, dynamical decoupling must be applied judiciously. We propose a software framework *Adaptive Dynamical Decoupling (ADAPT)* to identify the subset of qubits that provides the highest reliability with DD. ADAPT uses decoy circuits and a localized search algorithm to perform a trial-and-error search to identify the best subset of qubits to apply DD. We evaluate ADAPT on three quantum computers from IBM using two types of DD protocols and show that ADAPT improves fidelity by 1.86x compared to not using DD on average and by up-to 5.73x. Compared to DD on all qubits, ADAPT improves the fidelity of applications by 1.2x. | {'timestamp': '2021-09-14T02:14:08', 'yymm': '2109', 'arxiv_id': '2109.05309', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05309'} |
null | null |
# Introduction {#intro}
Let \(\gamma>1\). We consider the following Cauchy problem \[\label{general} \begin{split} \partial^{\alpha}_t(u-u_0)(t,x)+\Psi_{\beta}(-i\nabla)u(t,x) & =\abs{u(t,x)}^{\gamma-1}u(t,x),\quad t>0, \; x\in\mathbb{R}^d,\\ u(t,x)|_{t=0} & = u_0(x)\geq 0,\quad x\in\mathbb{R}^d, \end{split}\] where \(\partial^{\alpha}_t\) is the *Riemann--Liouville fractional derivative* of order \(\alpha\in(0,1)\) given by \[\partial^{\alpha}_t v=\dfrac{d}{dt}\displaystyle\int_0^t g_{1-\alpha}(t-s)v(s)ds=:\dfrac{d}{dt}(g_{1-\alpha}*v)(t),\] with \(g_{\rho}(t):=\frac{1}{\Gamma(\rho)}t^{\rho-1}\) and the Euler Gamma function \(\Gamma(\cdot)\). The function \(u_0\) stands for the initial data in a certain Lebesgue space. The term \(\Psi_{\beta}(-i\nabla)\) is a *singular integral operator* of constant order \(\beta\in(0,2)\) with symbol \(\psi\), that is \[\Psi_{\beta}(-i\nabla) v(x)= \mathcal{F}^{-1}_{\] where \(\mathcal{F}\) is the Fourier transform in \(\mathbb{R}^d\) and \(\mathcal{F}^{-1}\) is its inverse. As usual, \(C_0^{\infty}(\mathbb{R}^d)\) denotes the space of test functions on \(\mathbb{R}^d\). The symbol \(\psi\) is a measurable function on \(\mathbb{R}^d\) given by \[\psi(\] where \[\omega_\mu(\theta):=\displaystyle\int_{S^{d-1}}\abs{\theta\cdot\eta}^\beta\mu(d\eta),\quad\theta\in S^{d-1}.\] Here, \(\mu(d\eta)\) is a centrally symmetric finite (non-negative) Borel measure defined on the unitary sphere \(S^{d-1}\), called *spectral measure*, and \(\omega_\mu(\cdot)\) is a continuous function on \(S^{d-1}\), see e.g.. Whenever \(\mu(d\eta)=\varrho(\eta)d\eta\), where \(\varrho\) is a continuous function on \(S^{d-1}\), we will refer to \(\varrho\) as the *density of \(\mu\)*. Some restrictions on the function \(\varrho\) may be required for the lower bound and behaviour of the fundamental solutions; see, e.g.,. More precisely, our basic hypothesis throughout the paper is the following:
- The spectral measure \(\mu\) has a strictly positive density, such that the function \(\omega_\mu\) is strictly positive and \((d+1+[\beta])\)-times continuously differentiable on \(S^{d-1}\).
We denote by \((\mathcal{H}_2)\) to refer to \((\mathcal{H}_1)\) whenever we need to assume that \(\omega_\mu\) is \((d+2+[\beta])\)-times continuously differentiable on \(S^{d-1}\), \([\beta]\) being the maximal integer not exceeding the real number \(\beta\). The considerations just made above have been taken from and, for \(d=1\) and \(d>1\) respectively. We want to point out that the condition of strict positivity on \(\omega_\mu\) in \(\mathcal{H}_1\), guarantees that the support of the measure \(\mu\) on \(S^{d-1}\) is not contained in any hyperplane of \(\mathbb{R}^d\) (). In this work we are concerned with studying the blow-up phenomena for reaction-diffusion equations like [\[general\]](#general){reference-type="eqref" reference="general"} considering a non-regular class of solutions instead of the classical calls, with a temporal fractional derivative and a pseudo-differential operator related to a stochastic process. Our time derivative is also called the *Caputo fractional derivative*, for instance, in the sense of with \(0<\alpha<1\). However, other authors may require smoothness conditions on the function to define the Caputo derivative (see, e.g., ). Since \(\psi\) is a Lévy-Khintchine symbol with index of stability \(\beta\), it is well known that the corresponding stochastic process is called a localised Feller-Courrège process and therefore it makes sense to use the notation \(\Psi_{\beta}(-i\nabla)\) for the associated generator (see, e.g., ). In this setting, the Green function of the Cauchy problem \(\dfrac{\partial u}{\partial t}+\Psi_{\beta}(-i\nabla) u=0\) is non-Gaussian and it is interpreted as the transition probability density of the corresponding *stable non-Gaussian process*. The study of these processes and their generalizations is motivated by the increasing use in the mathematical modeling of processes in engineering, natural sciences and economics. See, e.g., and. Similar to the case of Gaussian processes, which have been widely studied (see, e.g.,, ,, , ), it arises an interest in qualitative properties, blow-up and asymptotic behavior for the solutions of non-Gaussian ones. For instance, in the case \(\beta=2\) and \(\omega_\mu\equiv 1\) we see that the operator, namely \(\Psi_2(-i\nabla)\), becomes the negative Laplacian \((-\Delta)\) with symbol \(\psi( \begin{equation*} \begin{split} \partial_t\;u(t,x)+(-\Delta)u(t,x) & = u(t,x)^{\gamma},\quad t>0, \; x\in\mathbb{R}^d,\\ u(t,x)|_{t=0} & = u_0(x)\geq 0,\quad x\in\mathbb{R}^d, \end{split} \end{equation*} was investigated by Fujita in 1966 (). Since then, many other researchers have explored blow-up phenomena (see, e.g.,, ,, ,, ). Following the analysis of this phenomenon, in this work we show that the non-linearity of \eqref{general} leads to the blow-up of positive solutions in a finite time. For this purpose, we say that a function\)u:\[0,T)\^d\(\textit{blows-up} at the finite time\)T\(if \[ \lim_{t\rightarrow T^-}\norm{u(t)}_\infty =+\infty, \] and thus our main result is stated as follows. \begin{theorem} \label{blowUp} Let\)(0,1)\(and\)(0,2)\(. Assume the hypothesis\)(\_1)\(holds. Suppose that\)=\(, that\)`<!-- -->`{=html}1\<p\<\(and that\)u_0L_p(\^d)C(\^d)\(is a non-negative function. If\)`<!-- -->`{=html}1\<\<1+ \(then all non-trivial non-negative solutions of \eqref{general} that admit the representation \eqref{integral} blow-up in finite time. Furthermore, if\)+ \(then the non-trivial non-negative solutions blow-up in finite time whenever the initial condition is sufficiently large. \end{theorem} For the proof, we exploit the representation of the Volterra equations in the sense of Prüss as well as the theory of \textit{completely positive kernels} of type (\[), like\)g\_\(with\)(0,1)\((see, e.g., ). We also deal with the theory of pseudo-differential operators that generate a sub-Markovian semigroup on\)L_p(\^d)\(, under the condition that the symbol\):\^d\(is a continuous and negative definite function; see. In our case,\]satisfies such condition (, ). Other results in and, are also particulary important for our work. Here, the authors show that the linear Cauchy problem \begin{equation*} \begin{split} \partial^{\alpha}_t(u-u_0)(t,x)+\Psi_{\beta}(-i\nabla)u(t,x)&=f(t,x),\quad t>0, \; x\in\mathbb{R}^d,\\ u(t,x)|_{t=0}&=u_0(x),\quad x\in\mathbb{R}^d, \end{split} \end{equation*} admits a pair of fundamental solutions\)(Z,Y)\(, given by \begin{equation}\label{Z:G} Z(t,x):=\dfrac{1}{\alpha}\displaystyle\int_0^\infty G(t^\alpha s,x)s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds \end{equation} and \begin{equation}\label{Y:G} Y(t,x):=\displaystyle\int_0^\infty t^{\alpha-1}G(t^\alpha s,x)s^{-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds, \end{equation} where\)G\(stands for the Green function that solves the problem \begin{equation}\label{HomogeneaOrden1} \begin{split} \partial_t\, G(t,x)+\Psi_{\beta}(-i\nabla)G(t,x)&=0,\quad t>0, \; x\in\mathbb{R}^d,\\ G(t,x)|_{t=0}&=\delta_0(x),\quad x\in\mathbb{R}^d, \end{split} \end{equation}\)\_0\(being the Dirac delta distribution, and\)G\_(,)\(is the Green function that solves the problem \[ \partial_t\, v(t,s)+\dfrac{d^\alpha}{ds^\alpha}v(t,s)=0,\quad t>0,\;s\in\mathbb{R}, \; G_\alpha(0,s)=\delta(s), \] where\)(0,1)\(and \[\dfrac{d^\alpha}{ds^\alpha}f(s):=\dfrac{1}{\Gamma(-\alpha)}\int_0^\infty \dfrac{f(s-\tau)-f(s)}{\tau^{1+\alpha}}d\tau, \] see. From, we know that the fundamental solutions\)(Z,Y)\(given by \eqref{Z:G}-\eqref{Y:G} satisfy the relation \begin{equation} \label{relacionYZ} Y(\cdot,x)=\dfrac{d}{dt}(g_{\alpha}*Z(\cdot,x)),\quad t>0,\quad x\in\mathbb{R}^d\setminus\{0\}, \end{equation} which is crucial for the integral representation \eqref{integral} below. This paper is organized as follows. In Section \ref{sec:1} we have compiled some properties of the fundamental solutions\)Z,Y\(. In Section \ref{sec:2} we derive the integral representation of a solution to \eqref{general} in the sense of Definition \ref{defSolution}. The main result of this section is given by Theorem \ref{localrepresentation}. In Section \ref{sec:3} we prove two results on continuity and non-negativeness of the local solutions, given by Theorems \ref{continuidad} and \ref{positividad} respectively. The last section is devoted to prove our main result related to the blow-up of positive solutions, stated in Theorem \ref{blowUp}. We also give a Fujita type result for global solutions in Theorem \ref{SolucionGlobal}. \section{Preliminaries} \label{sec:1} In what follows we use the notations\)fg\(and\)fg\(in\)D\(, which means that there exists constants\)C,C_1,C_2\>0\(such that\)C_1 gfC_2 g\(and\)fC g\(in\)D\(, respectively. Such constants may change line by line. We also use the notation\)=\^t\^-\(for\)x\^d\(and\)t\>0\(. \medbreak The following result summarize the two-side estimates for\)Z\(. For its proof see . \begin{proposition} \label{cotasZ} Let\)(0,1)\(and\)(0,2)\(. Assume the hypothesis\)(\_1)\(holds. Then there exists a positive constant\)C\(such that for\)(t,x)(0,)\^d\(the following two-sided estimates for\)Z\(hold. For\[, \begin{align*} Z(t,x)\asymp C t^{-\frac{\alpha d}{\beta}}~~~~~~~~~~~~~~~&\quad\text{if}\quad d<\beta,\\ Z(t,x)\asymp C t^{-\alpha}(|\log(\Omega)|+1) &\quad\text{if}\quad d=\beta,\\ Z(t,x)\asymp C t^{-\frac{\alpha d}{\beta}}\Omega^{1-\frac{d}{\beta}}~~~~~~~~ &\quad\text{if}\quad d>\beta. \end{align*} For\], \begin{equation*} Z(t,x)\asymp C t^{-\frac{\alpha d}{\beta}}\Omega^{-1-\frac{d}{\beta}}. \end{equation*} \end{proposition} In the same way we have derived the two-side estimates for\)Y\(. \begin{proposition} \label{cotasY} Under the same assumptions as Proposition \ref{cotasZ}, the following two-sided estimates for\)Y\(hold. For\[, \begin{align*} Y(t,x)\asymp C t^{-\frac{\alpha d}{\beta}+\alpha-1}~~~~~~~~~~~~\,&\quad \text{if}\quad d<2\beta,\\ Y(t,x)\asymp C t^{-\alpha-1}(|\log(\Omega)|+1)~ &\quad \text{if}\quad d=2\beta, \\ Y(t,x)\asymp C t^{-\frac{\alpha d}{\beta}+\alpha-1}\Omega^{2-\frac{d}{\beta}}~~~~~~&\quad \text{if}\quad d>2\beta. \end{align*} For\], \begin{equation*} Y(t,x)\asymp C t^{-\frac{\alpha d}{\beta}+\alpha-1}\Omega^{-1-\frac{d}{\beta}}. \end{equation*} \end{proposition} \begin{proof} The assertions follow from a straightforward computations made in the proof of the previous estimates for\)Z\(. \end{proof} \begin{remark} We note a singularity at the origin with respect to the spatial variable for\)Z\(, whenever\)d\(, and for\)Y\(whenever\)d\(. It is well known that this type of singularities occurs in the equations of fractional evolution in time, even if\[and\)\_\(. \end{remark} Now we continue by showing some properties of the fundamental solutions\)Z\(and\)Y\(. \begin{lemma} \label{cotasDeltaZ} Under the same assumptions as Proposition \ref{cotasZ}, there exists a positive constant\)C\(for all\)t_1,t_2\>0\(and\)x\^d\(, such that there exists\)t_c\>0\(, between\)t_1\(and\)t_2\(, and the following estimates for\)Z\(hold with\)\_c=\^t_c\^-\(. For\)\_c\(, \begin{equation*} |Z(t_1,x)-Z(t_2,x)|\leq C |t_1-t_2| \begin{cases} t_c^{-\frac{\alpha d}{\beta}-1} &\quad \text{if}\quad d<\beta,\\ t_c^{-\alpha-1}(|\log(\Omega_c)|+1) &\quad \text{if}\quad d=\beta,\\ t_c^{-\frac{\alpha d}{\beta}-1}\Omega_c^{1-\frac{d}{\beta}}&\quad \text{if}\quad d>\beta, \end{cases} \end{equation*} and for\)\_c\(, \[ |Z(t_1,x)-Z(t_2,x)|\leq C |t_1-t_2| t_c^{-\frac{\alpha d}{\beta}-1}\Omega_c^{-1-\frac{d}{\beta}}. \] \end{lemma} \begin{proof} From \eqref{Z:G} it follows that \[ Z(t_1,x)-Z(t_2,x)=\dfrac{1}{\alpha}\displaystyle\int_0^\infty \left[G(t_1^\alpha s,x)-G(t_2^\alpha s,x)\right] s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds. \] It is known () that\)G\(is differentiable with respect to\)t\>0\(and satisfies \begin{equation*} \begin{split} |G(t,x)|\leq C \min\left(t^{-\frac{d}{\beta}},\dfrac{t}{\norm{x}^{d+\beta}}\right),\\ \left|t\dfrac{\partial G}{\partial t}(t,x)\right |\leq C \min\left(t^{-\frac{d}{\beta}},\dfrac{t}{\norm{x}^{d+\beta}}\right). \end{split} \end{equation*} In these estimates,\)C\(depends on\],\)d\(and the bounds for\)\_\(. Using this and the mean-value theorem, we have that for some\)t_c\(between\)t_1\(and\)t_2\(, \begin{equation*} \begin{split} |Z(t_1,x)-Z(t_2,x)|&\leq \dfrac{1}{\alpha}\displaystyle\int_0^\infty \left|\dfrac{\partial G(t^\alpha s,x)}{\partial t}\right|_{t=t_c}|t_1-t_2| s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds\\ &\leq \dfrac{|t_1-t_2|}{t_c}\displaystyle\int_0^\infty \left|t_c^\alpha s G'(t_c^\alpha s,x)\right| s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds\\ &\leq C \dfrac{|t_1-t_2|}{t_c}\displaystyle\int_0^\infty \min\left((t_c^\alpha s)^{-\frac{d}{\beta}}, \dfrac{t_c^\alpha s}{\norm{x}^{d+\beta}}\right) s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds. \end{split} \end{equation*} From and one finds that the asymptotic behavior of\)G\_\(is the same as for the density\)w\_\(given in (with the skewness of the distribution that equals to\)`<!-- -->`{=html}0\(). Therefore, we proceed in the same way from, i.e., \begin{align*} &\displaystyle\int_0^\infty \min\left((t_c^\alpha s)^{-\frac{d}{\beta}}, \dfrac{t_c^\alpha s}{\norm{x}^{d+\beta}}\right) s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds\\ &=\displaystyle\int_0^1 \min\left((t_c^\alpha s)^{-\frac{d}{\beta}}, \dfrac{t_c^\alpha s}{\norm{x}^{d+\beta}}\right) s^{-1-\frac{1}{\alpha}}w_\alpha(s^{-\frac{1}{\alpha}})ds+\displaystyle\int_1^\infty \min\left((t_c^\alpha s)^{-\frac{d}{\beta}}, \dfrac{t_c^\alpha s}{\norm{x}^{d+\beta}}\right) s^{-1-\frac{1}{\alpha}}w_\alpha(s^{-\frac{1}{\alpha}})ds. \end{align*} Thus, we get the desired bounds. \end{proof} \begin{lemma} \label{cotasDeltaY} Under the same assumptions as Proposition \ref{cotasZ}, there exists a positive constant\)C\(for all\)t_1,t_2\>0\(and\)x\^d\(, such that there exists\)t_c\>0\(, between\)t_1\(and\)t_2\(, and the following estimates for\)Y\(hold with\)\_c=\^t_c\^-\(. For\)\_c\(, \[ |Y(t_1,x)-Y(t_2,x)|\leq C |t_1-t_2| \begin{cases} t_c^{-\frac{\alpha d}{\beta}+\alpha-2} &\quad \text{if}\quad d<2\beta,\\ t_c^{-\alpha-2}(|\log(\Omega_c)|+1) &\quad \text{if}\quad d=2\beta,\\ t_c^{-\frac{\alpha d}{\beta}+\alpha-2}\Omega_c^{2-\frac{d}{\beta}}&\quad \text{if}\quad d>2\beta, \end{cases} \] and for\)\_c\(, \[ |Y(t_1,x)-Y(t_2,x)|\leq C |t_1-t_2| t_c^{-\frac{\alpha d}{\beta}+\alpha-2}\Omega_c^{-1-\frac{d}{\beta}}. \] \end{lemma} \begin{proof} The assertions follow from a straightforward computations made in the proof of the previous estimates for\)Z\(, but using \eqref{Y:G}. \end{proof} \medbreak \begin{lemma} \label{cotasDeltaEspacialZ} Let\)(0,1)\(and\)(0,2)\(. Assume the hypothesis\)(\_2)\(holds. Then there exists a positive constant\)C\(for all\)t\>0\(and\)x_1,x_2\^d\(, such that there exists\[in the open segment connecting\)x_1\(and\)x_2\(, and the following estimates for\)Z\(hold with\)\_=\^t\^-\(. For\)\_\(, \[ |Z(t,x_1)-Z(t,x_2)|\leq C \norm{x_1-x_2}t^{-\frac{\alpha (d+1)}{\beta}}\Omega_\zeta^{1-\frac{d+1}{\beta}} \] and for\)\_\(, \[ |Z(t,x_1)-Z(t,x_2)|\leq C \norm{x_1-x_2} t^{-\frac{\alpha (d+1)}{\beta}}\Omega_\zeta^{-1-\frac{d+1}{\beta}}. \] \end{lemma} \begin{proof} From \eqref{Z:G} it follows that \[ Z(t,x_1)-Z(t,x_2)=\dfrac{1}{\alpha}\displaystyle\int_0^\infty \left[G(t^\alpha s,x_1)-G(t^\alpha s,x_2)\right] s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds. \] From () we know that\)G\(is one time continuously differentiable in\)x\(and satisfies, for any\)j=1,, d\(, \begin{equation*} \left|\dfrac{\partial G}{\partial x_j}(t,x)\right |\leq C \min\left(t^{-\frac{d+1}{\beta}},\dfrac{t}{\norm{x}^{d+\beta+1}}\right). \end{equation*} We recall that\)C\(depends on\],\)d\(and the bounds for\)\_\(. Let\)DG(t\^s,x)\(the Jacobian of\)G(t\^s,)\(in the point\)x\(. By the mean-value inequality, we have that for some\[in the open segment between\)x_1\(and\)x_2\(, \begin{equation*} \begin{split} |Z(t,x_1)-Z(t,x_2)|&\leq \dfrac{1}{\alpha}\displaystyle\int_0^\infty \abs{DG(t^\alpha s,\zeta)(x_1-x_2)} s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds\\ &\leq \dfrac{\norm{x_1-x_2}}{\alpha}\displaystyle\int_0^\infty \sum_{j=1}^d\left|\dfrac{\partial G}{\partial x_j}(t^\alpha s,\zeta)\right| s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds\\ &\lesssim\norm{x_1-x_2}\displaystyle\int_0^\infty \min\left((t^\alpha s)^{-\frac{d+1}{\beta}}, \dfrac{t^\alpha s}{\norm{\zeta}^{d+\beta+1}}\right) s^{-1-\frac{1}{\alpha}}G_\alpha(1,s^{-\frac{1}{\alpha}})ds. \end{split} \end{equation*} Now, we proceed in the same way from. \end{proof} \begin{lemma} \label{cotasDeltaEspacialY} Under the same assumptions as Lemma \ref{cotasDeltaEspacialZ}, then there exists a positive constant\)C\(for all\)t\>0\(and\)x_1,x_2\^d\(, such that there exists\]in the open segment connecting\)x_1\(and\)x_2\(, and the following estimates for\)Y\(hold with\)\_=\^t\^-\(. For\)\_\(, \begin{equation*} |Y(t,x_1)-Y(t,x_2)|\leq C \norm{x_1-x_2} \begin{cases} t^{-\frac{\alpha (d+1)}{\beta}+\alpha-1} &\quad \text{if}\quad d+1<2\beta,\\ t^{-\alpha-1}(|\log(\Omega_\zeta)|+1) &\quad \text{if}\quad d+1=2\beta,\\ t^{-\frac{\alpha (d+1)}{\beta}+\alpha-1}\Omega_\zeta^{2-\frac{d+1}{\beta}}&\quad \text{if}\quad d+1>2\beta, \end{cases} \end{equation*} and for\)\_\(, \[ |Y(t,x_1)-Y(t,x_2)|\leq C \norm{x_1-x_2} t^{-\frac{\alpha (d+1)}{\beta}+\alpha-1}\Omega_\zeta^{-1-\frac{d+1}{\beta}}. \] \end{lemma} \begin{proof} This is similar to the proof of the previous lemma for\)Z\(. \end{proof} \begin{lemma} \label{NormaGradienteY} Under the same assumptions as Lemma \ref{cotasDeltaEspacialZ}, then there exists a positive constant\)C\(for all\)t\>0\(and\)x_1,x_2\^d\(, such that the estimate \begin{equation} \label{cotasNablaY} \norm{Y(t,\cdot-x_1)-Y(t,\cdot-x_2)}_q\leq C \norm{x_1-x_2}\norm{\nabla Y(t,\cdot)}_q\lesssim \norm{x_1-x_2} t^{-\frac{\alpha d}{\beta}\left(1-\frac{1}{q}\right)-\frac{\alpha}{\beta}+\alpha-1} \end{equation} is true for\)`<!-- -->`{=html}1q\<'\(, where\)':=
\(. In the case of\)d+1\<2\(, \eqref{cotasNablaY} remains true for\)q=\(. \end{lemma} \begin{proof} It follows from the same arguments as in but using the bounds given in Lemma \ref{cotasDeltaEspacialY}. \end{proof} \begin{definition} \label{defSolution} Let\)(0,1)\(,\)(0,2)\(and\)\>1\(. Assume the hypothesis\)(\_1)\(holds. Suppose that\)`<!-- -->`{=html}1\<p\<\(and that\)u_0L_p(\^d)\(is a non-negative function. A function\)u\(is called a \textbf{local solution} of \eqref{general} if there exists\)T\>0\(sucht that \begin{enumerate}[(i)] \item\)uC(\[0,T\];L_p(\^d))L\_((0,T)\^d)\(, \item\)u\(satisfies \eqref{general} in\)\[0,T\]\(. \end{enumerate} A function\)u\(is called a \textbf{global solution} of \eqref{general} if\)(i)\(-\)(ii)\(are satisfied for any\)T\>0\(. \end{definition} At this point, we mention that problems like \eqref{general} have been studied in. Under suitable conditions on\[,\],\[and\)p\(, together with other parameters, the authors find positive local and global solutions. \section{Representation of solution in its integral form} \label{sec:2} In this section we analyse the conditions under which a local solution\)u\(of \eqref{general}, in the sense of Definition \ref{defSolution}, can be represented as \begin{equation} \label{integral} u(t,x) = \int_{\mathbb{R}^d}Z(t,x-y)u_0(y)dy + \int_0^t \int_{\mathbb{R}^d}Y(t-s,x-y)|u(s,y)|^{\gamma-1}u(s,y)dyds \end{equation} for all\)x\^d\(and\)`<!-- -->`{=html}0t\< T\(. First, we recall that the symbol\)( \]\int_{\mathbb{R}^d}\left(-\Psi_{\beta}(-i\nabla)f\right)(x)\left((f-1)^+\right)^{p-1}(x)dx\leq 0,\quad f\in C_0^{\infty}(\mathbb{R}^d),\[ and consequently it is \(L_p(\mathbb{R}^d)\)-dissipative (). In fact, the density of \(C_0^{\infty}(\mathbb{R}^d)\) in \(L_p(\mathbb{R}^d)\) implies that \((-\Psi_{\beta}(-i\nabla), C_0^{\infty}(\mathbb{R}^d))\) is closable and its closure \((A, D(A))\) generates a sub-Markovian semigroup \(\{T_t\}_{t\geq 0}\) on \(L_p(\mathbb{R}^d)\) which is a strongly continuous contraction semigroup (). Besides, \(A\) is densely defined on \(L_p(\mathbb{R}^d)\) (). On the other hand, it is well known that \(g_\alpha\) is a completely positive function and belongs to \(L_{1,loc}(\mathbb{R}^+)\). In what follows, we denote \(u(t)=u(t,\cdot)\) and \(|u|^{\gamma-1}(t)=\abs{u(t,\cdot)}^{\gamma-1}\). Since \(u\) and \(u_0\) satisfy Definition [\[defSolution\]](#defSolution){reference-type="ref" reference="defSolution"}, if \(u_0\in L_\infty(\mathbb{R}^d)\) we observe that \(g_{\alpha}*|u|^{\gamma-1}u(t)\in L_p(\mathbb{R}^d)\) for \(0\leq t<T\) and the equation in [\[general\]](#general){reference-type="eqref" reference="general"} can be written as the Volterra equation \]\label{Volterra} u(t)=u_0+g_{\alpha}*|u|^{\gamma-1}u(t)+g_{\alpha}*Au(t), \quad 0<t<T,\[ which admits a resolvent \(\{S(t)\}_{t\geq 0}\) in \(L_p(\mathbb{R}^d)\) (). From we have that \]S(t)=-\int_0^\infty T_\tau \; w(t;d\tau),\quad t>0,\[ where \(w\) is the propagation function associated with \(g_\alpha\). In order to describe this resolvent, we use the representation \]T_tf(\cdot)=\int_{\mathbb{R}^d} G(t,\cdot-y)f(y)dy, \quad f\in D(A),\[ the function \(G\) being the fundamental solution of the problem [\[HomogeneaOrden1\]](#HomogeneaOrden1){reference-type="eqref" reference="HomogeneaOrden1"} (see ). For \(v\in D(A)\) we see that \]\begin{aligned}
S(t)v&=-\int_0^\infty T_\tau v\; w(t;d\tau)\\ &=-\int_0^\infty G(t,\cdot)\star v\; w(t;d\tau) \end{aligned}\[ and using the Fourier transform we obtain \]\begin{aligned}
\mathcal{F}(S(t)v)&=-\int_0^\infty e^{-\tau\psi( &=s(t,\psi( &=\widehat{Z}(t, \end{aligned}\[ with a kernel \(s\) that comes via scalar Volterra equations (see, ). This implies that \]S(t)v=Z(t,\cdot)\star v\[ and the boundedness of \(S(t)\) leads to an extension to all of \(L_p(\mathbb{R}^d)\). Let \(0<t<T\). If \(u(s)\in D(A)\), \(0\leq s\leq t\), identity [\[Volterra\]](#Volterra){reference-type="eqref" reference="Volterra"} and yield \]\begin{aligned}
1*u(t)&=\int_0^t u(s)ds\\ &=\int_0^t \left(S(t-s)u(s)-A(g_\alpha * S)(t-s)u(s)\right) ds\\ &=\int_0^t S(t-s)u(s)ds-\int_0^t(g_\alpha * S)(t-s)Au(s)ds\\ &=\int_0^t S(s)u(t-s)ds-\int_0^t S(s)(g_\alpha * Au)(t-s) ds\\ &=\int_0^t S(s)\left(u(t-s)-(g_\alpha * Au)(t-s)\right) ds\\ &=\int_0^t S(s)\left(u_0+g_{\alpha}*|u|^{\gamma-1}u(t-s)\right) ds \end{aligned}\[ and thus we get the *variation of parameters formula* for [\[Volterra\]](#Volterra){reference-type="eqref" reference="Volterra"} given by \]u(t)=\frac{d}{dt}\int_0^t S(s)\left(u_0+g_{\alpha}*|u|^{\gamma-1}u\right)(t-s) ds.\[ We note that \]\frac{d}{dt}\int_0^t S(s)u_0ds=S(t)u_0=Z(t,\cdot)\star u_0.\[ By proceeding as in the proof of, but working with the \(L_p(\mathbb{R}^d)\) space, using the relation [\[relacionYZ\]](#relacionYZ){reference-type="eqref" reference="relacionYZ"} and the fact that \(\underset{0\leq t<T}{\sup}{\left\||u|^{\gamma-1}u(t)\right\|}_\infty<\infty\), we show that \]\frac{d}{dt}\int_0^t S(s)\left(g_{\alpha}*|u|^{\gamma-1}u\right)(t-s)ds=\int_0^t Y(t-s,\cdot)\star|u|^{\gamma-1}u(s,\cdot)ds.\[
# Continuity and non-negativeness of solution in \([0,T)\times \mathbb{R}^d\) {#sec:3}
Let \(u\) be a local solution of [\[general\]](#general){reference-type="eqref" reference="general"}. In this section we show that \(u\) is a continuous and non-negative function on \([0,T)\times \mathbb{R}^d\), for some \(T>0\). For this purpose, the representation [\[integral\]](#integral){reference-type="eqref" reference="integral"} obtained in the previous section is particularly important. In what follows we use the parameter \(\kappa:=\begin{cases}\frac{d}{\beta}, & d>\beta,\\ 1, & otherwise \end{cases}\) which sets a condition on \(p\) for the existence of some \(q\geq 1\) such that \]\frac{1}{p}+\frac{1}{q}=1\[ and thus we can reach the \(L_q\)-norm for \(Y(t,\cdot)\), \(t>0\) (see ).
The second result of this section is the following.
# Proof of the main result Theorem [\[blowUp\]](#blowUp){reference-type="ref" reference="blowUp"} {#sec:4}
Firstly, we get the following estimates. Let \(t>0\). Using the bounds given in Proposition [\[cotasZ\]](#cotasZ){reference-type="ref" reference="cotasZ"}, it is clear that \]Z(t,x-y)\geq C t^{-\frac{\alpha d}{\beta}} e^{-\frac{\norm{x-y}^2}{4t}},\quad \Omega\leq 1.\[ If \(\Omega\geq 1\), we have that \]\begin{aligned}
Z(t,x-y)&\geq C t^{-\frac{\alpha d}{\beta}}\Omega^{-1-\frac{d}{\beta}}\\ &= C t^{-\frac{\alpha d}{\beta}}t^{\alpha+\frac{\alpha d}{\beta}}\norm{x-y}^{-\beta-d}\\ &= C t^{-\frac{\alpha d}{\beta}}t^{\alpha+\frac{\alpha d}{\beta}}(2\sqrt{t})^{-\beta-d}\left(\frac{\norm{x-y}}{2\sqrt{t}}\right)^{-\beta-d}\\ &\geq C t^{-\frac{\alpha d}{\beta}}t^{\alpha+\frac{\alpha d}{\beta}}(2\sqrt{t})^{-\beta-d}e^{-\frac{\norm{x-y}^2}{4t}}, \end{aligned}\[ whenever \(d\leq 3\). For larger dimensions, it is always possible to find a suitable constant \(K>1\), depending on \(\beta\) and \(d\), such that \(\left(\frac{\norm{x-y}}{2\sqrt{t}}\right)^{-\beta-d}\geq e^{-K\frac{\norm{x-y}^2}{4t}}\). From the hypothesis \(\alpha=\frac{\beta}{2}\), it follows that \]Z(t,x-y)\geq C 2^{-\beta-d} t^{-\frac{\alpha d}{\beta}} e^{-\frac{\norm{x-y}^2}{4t}},\quad \Omega\geq 1,\[ which means that \]\label{Zvscampana} Z(t,x-y)\geq C_1 t^{-\frac{\alpha d}{\beta}} e^{-\frac{\norm{x-y}^2}{4t}},\[ for all \(t>0\) and \(x,y\in\mathbb{R}^d\), with \(C_1=\dfrac{C}{2^{\beta+d}}\). We may assume without loss of generality that the constant \(C\) of the Proposition [\[cotasY\]](#cotasY){reference-type="ref" reference="cotasY"} is the same as that of the Proposition [\[cotasZ\]](#cotasZ){reference-type="ref" reference="cotasZ"}. In this way, we have also derived \]\label{Yvscampana} Y(t-s,x-y)\geq C_1 (t-s)^{-\frac{\alpha d}{\beta}+\alpha-1} e^{-\frac{\norm{x-y}^2}{4(t-s)}},\[ for all \(0\leq s < t\) and \(x,y\in\mathbb{R}^d\). Now, we proceed by contradiction. We suppose that there exists a global non-trivial solution \(u\) of [\[general\]](#general){reference-type="eqref" reference="general"}, according to Definition [\[defSolution\]](#defSolution){reference-type="ref" reference="defSolution"}. In this case, \(u_0(y_0)>0\) for some \(y_0\in\mathbb{R}^d\). The continuity of \(u_0\) implies that \]u_0(y)> C_0, \quad \forall y\in B(y_0,\delta),\[ with some \(\delta >0\) and \(C_0=\dfrac{u_0(y_0)}{2}\). The representation [\[integral\]](#integral){reference-type="eqref" reference="integral"} for \(u\) is \]u(t,x) = \int_{\mathbb{R}^d}Z(t,x-y)u_0(y)dy + \int_0^t \int_{\mathbb{R}^d}Y(t-s,x-y)u(s,y)^{\gamma}dyds\[ for all \(x\in\mathbb{R}^d\) and \(0<t< T\). We note that, given the assumption made, \(T\) can be arbitrarily large. As in Section [2](#sec:3){reference-type="ref" reference="sec:3"}, we define \]u_1(t,x):= \int_{\mathbb{R}^d}Z(t,x-y)u_0(y)dy\[ and \]u_2(t,x):=\int_0^t \int_{\mathbb{R}^d}Y(t-s,x-y)u(s,y)^{\gamma}dyds.\[ Using [\[Zvscampana\]](#Zvscampana){reference-type="eqref" reference="Zvscampana"}, it follows that \]\begin{aligned}
u_1(t,x)&\geq C_1 t^{-\frac{\alpha d}{\beta}} \int_{\mathbb{R}^d}e^{-\frac{\norm{x-y}^2}{4t}} u_0(y)dy\\ &\geq C_1C_0 t^{-\frac{\alpha d}{\beta}}\int_{B(y_0,\delta)}e^{-\frac{\norm{x-y}^2}{4t}}dy\\ &\geq C_1C_0 t^{-\frac{\alpha d}{\beta}}e^{-\frac{\norm{x-y_0}^2}{2t}}\int_{B(y_0,\delta)}e^{-\frac{\norm{y-y_0}^2}{2t}}dy \end{aligned}\[ and we obtain \]\label{EstiPartLineal} u_1(t,x)\geq C_2 t^{-\frac{\alpha d}{\beta}}e^{-\frac{\norm{x}^2}{t}}, \quad t>1,\quad x\in\mathbb{R}^d.\[ Let \(H\) be the heat kernel \]H(t,x)=\frac{1}{(4\pi t)^{\frac{d}{2}}}e^{-\frac{\norm{x}^2}{4t}},\quad t>0,\quad x\in\mathbb{R}^d.\[ Using the fact that \]\int_{\mathbb{R}^d}H(t,x)dx=1,\[ we define the function \]\label{F} F(t)=\int_{\mathbb{R}^d}H(t,x)u(t,x)dx,\quad t>0,\[ and splitting the integral into two parts we see that \]F(t)=\int_{\mathbb{R}^d}H(t,x)u_1(t,x)dx+\int_{\mathbb{R}^d}H(t,x)u_2(t,x)dx.\[ In the first integral we use the estimate [\[EstiPartLineal\]](#EstiPartLineal){reference-type="eqref" reference="EstiPartLineal"}, for obtaining \]F(t)\geq C_3 t^{-\frac{\alpha d}{\beta}}+\int_{\mathbb{R}^d}H(t,x)u_2(t,x)dx\[ whenever \(t>1\). In the second integral, we use the fact that (see ) \]\frac{1}{g_\alpha(t)}\displaystyle\int_{\mathbb{R}^d}Y(t,x)dx=1,\quad t>0.\[ Jensen's inequality and Fubini's theorem yield \]\begin{aligned}
&\int_{\mathbb{R}^d}H(t,x)u_2(t,x)dx\\ &=\int_{\mathbb{R}^d}H(t,x)\left[\int_0^t \int_{\mathbb{R}^d}Y(t-s,x-y)u(s,y)^{\gamma}dy ds\right] dx\\ &=\int_0^t g_\alpha(t-s)\int_{\mathbb{R}^d}H(t,x)\left[\int_{\mathbb{R}^d}\frac{1}{g_\alpha(t-s)} Y(t-s,x-y)u(s,y)^{\gamma}dy\right] dx ds\\ &\geq \int_0^t g_\alpha(t-s)\int_{\mathbb{R}^d}H(t,x)\left[\int_{\mathbb{R}^d} \frac{1}{g_\alpha(t-s)} Y(t-s,x-y)u(s,y)dy\right]^{\gamma} dx ds\\ &= \int_0^t \left(g_\alpha(t-s)\right)^{1-\gamma}\int_{\mathbb{R}^d}H(t,x)\left[\int_{\mathbb{R}^d}Y(t-s,x-y)u(s,y)dy\right]^{\gamma} dx ds\\ &\geq \int_0^t \left(g_\alpha(t-s)\right)^{1-\gamma}\left[\int_{\mathbb{R}^d}H(t,x)\int_{\mathbb{R}^d}Y(t-s,x-y)u(s,y)dy\; dx\right]^{\gamma} ds\\ &\geq \int_0^t \left(g_\alpha(t-s)\right)^{1-\gamma}\left\{\int_{\mathbb{R}^d}\left[\int_{\mathbb{R}^d}H(t,x)Y(t-s,x-y)dx \right] u(s,y)dy\right\}^{\gamma} ds. \end{aligned}\[ The expression in the square brackets can be estimated with [\[Yvscampana\]](#Yvscampana){reference-type="eqref" reference="Yvscampana"}, i.e., \]\begin{aligned}
&\int_{\mathbb{R}^d}H(t,x)Y(t-s,x-y)dx\\ &\geq C_1 (t-s)^{-\frac{\alpha d}{\beta}+\alpha-1} \int_{\mathbb{R}^d}H(t,x)e^{-\frac{\norm{x-y}^2}{4(t-s)}}dx\\ &= C_1 (4\pi s)^{-\frac{d}{2}}e^{-\frac{\norm{y}^2}{4s}}\left(\frac{s}{t}\right)^{\frac{d}{2}} (t-s)^{-\frac{\alpha d}{\beta}+\alpha-1} \int_{\mathbb{R}^d}e^{\frac{\norm{y}^2}{4s}-\frac{\norm{x}^2}{4t}-\frac{\norm{x-y}^2}{4(t-s)}}dx. \end{aligned}\[ Proceeding in the same way as in, with \(\alpha=\frac{\beta}{2}\), we get \]\int_{\mathbb{R}^d}H(t,x)Y(t-s,x-y)dx \geq C_4 (4\pi s)^{-\frac{d}{2}}e^{-\frac{\norm{y}^2}{4s}}\left(\frac{s}{t}\right)^{\frac{d}{2}} (t-s)^{\alpha-1}\[ and thus \]\begin{aligned}
&\left\{\int_{\mathbb{R}^d}\left[\int_{\mathbb{R}^d}H(t,x)Y(t-s,x-y)dx \right] u(s,y)dy\right\}^{\gamma}\\ &\geq C_4^\gamma (t-s)^{(\alpha-1)\gamma} \left(\frac{s}{t}\right)^{\frac{d}{2}\gamma}\left\{\int_{\mathbb{R}^d}(4\pi s)^{-\frac{d}{2}}e^{-\frac{\norm{y}^2}{4s}}u(s,y)dy\right\}^\gamma\\ &=C_4^\gamma(t-s)^{(\alpha-1)\gamma} \left(\frac{s}{t}\right)^{\frac{d}{2}\gamma}F^\gamma(s) \end{aligned}\[ for \(0<s<t\). It follows that \]\int_{\mathbb{R}^d}H(t,x)u_2(t,x)dx\geq C_4^\gamma\int_0^t \left(g_\alpha(t-s)\right)^{1-\gamma}(t-s)^{(\alpha-1)\gamma} \left(\frac{s}{t}\right)^{\frac{d}{2}\gamma}F^\gamma(s) ds\[ and hence \]F(t) \geq \frac{C_3}{t^{\frac{d}{2}}}+C_5 \frac{t^{\alpha-1}}{t^{\frac{d}{2}\gamma}}\int_0^t s^{\frac{d}{2}\gamma}F^\gamma(s) ds\[ for all \(t>1\). Consequently, \]\label{FacotadaCF} t^{\frac{d}{2}\gamma}t^{1-\alpha} F(t) \geq C_3 t^{\frac{d}{2}(\gamma-1)}t^{1-\alpha}+C_5 \int_0^t s^{\frac{d}{2}\gamma}F^\gamma(s) ds.\[ Defining the r.h.s. of this expression as \(f(t)\), \(t>1\), we have that \]\label{cotaf} f(t)\geq C_3 t^{\frac{d}{2}(\gamma-1)}t^{1-\alpha}\[ and that \]\label{cotaf'} f'(t)\geq C_5 t^{\frac{d}{2}\gamma}F^\gamma(t).\[ From [\[FacotadaCF\]](#FacotadaCF){reference-type="eqref" reference="FacotadaCF"} it follows that \]\begin{aligned}
f'(t)&\geq C_5 t^{\frac{d}{2}\gamma}\left(\frac{f(t)}{t^{\frac{d}{2}\gamma+1-\alpha}}\right)^\gamma\\ &=C_5 t^{\frac{d}{2}\gamma(1-\gamma)-(1-\alpha)\gamma}f^\gamma(t). \end{aligned}\[ Therefore, \]f'(t)f^{-\gamma}(t)\geq C_5 t^{\frac{d}{2}\gamma(1-\gamma)-(1-\alpha)\gamma}\[ and \]\int_t^T f'(s)f^{-\gamma}(s)ds \geq C_5\int_t^T s^{\frac{d}{2}\gamma(1-\gamma)-(1-\alpha)\gamma}ds\[ with \(T>t\). From here, we get that \]\frac{f^{1-\gamma}(t)}{\gamma-1}\geq C_5\int_t^T s^{\frac{d}{2}\gamma(1-\gamma)-(1-\alpha)\gamma}ds\[ and using [\[cotaf\]](#cotaf){reference-type="eqref" reference="cotaf"} we also obtain the estimate \]\frac{f^{1-\gamma}(t)}{\gamma-1}\leq\frac{C_3^{1-\gamma}}{\gamma-1}t^{-\frac{d}{2}(1-\gamma)^2-(1-\alpha)(\gamma-1)}.\[ This implies that \]\label{exponentes} \frac{C_3^{1-\gamma}}{\gamma-1}t^{-\frac{d}{2}(1-\gamma)^2-(1-\alpha)(\gamma-1)}\geq C_5\int_t^T s^{-\frac{d}{2}\gamma(\gamma-1)-(1-\alpha)\gamma}ds.\[ Next we analyse the r.h.s. of [\[exponentes\]](#exponentes){reference-type="eqref" reference="exponentes"}, according to the following cases with \(a:=d-2(1-\alpha)\).\
For the case \(1<\gamma\leq \frac{a}{d}+\frac{2}{d\gamma}\), we have \]\begin{aligned}
\gamma\leq \frac{a}{d}+\frac{2}{d\gamma}&\Rightarrow d\gamma^2\leq a\gamma+2\\ &\Leftrightarrow d\gamma^2+2(1-\alpha)\gamma-d\gamma-2\leq 0\\ &\Leftrightarrow-\frac{d\gamma}{2}(\gamma-1)-(1-\alpha)\gamma+1\geq 0, \end{aligned}\[ which yields a contradiction for large enough \(T\). For the case \(\frac{a}{d}+\frac{2}{d\gamma}<\gamma< \frac{a}{d}+\frac{2}{d}\), we write the expression [\[exponentes\]](#exponentes){reference-type="eqref" reference="exponentes"} as \]\frac{C_3^{1-\gamma}}{\gamma-1}t^{-\frac{d}{2}(1-\gamma)^2-(1-\alpha)(\gamma-1)}\geq C_5\frac{t^{-\frac{d}{2}\gamma(\gamma-1)-(1-\alpha)\gamma+1}-T^{-\frac{d}{2}\gamma(\gamma-1)-(1-\alpha)\gamma+1}}{\frac{d}{2}\gamma(\gamma-1)+(1-\alpha)\gamma-1}.\[ Besides \]\begin{aligned}
\gamma<\frac{a}{d}+\frac{2}{d}&\Rightarrow d\gamma<d-2(1-\alpha)+2\\ &\Leftrightarrow-1<-\frac{d}{2}(\gamma-1)-(1-\alpha)\\ &\Leftrightarrow \frac{d\gamma}{2}(\gamma-1)+(1-\alpha)\gamma-1<\frac{d}{2}(\gamma-1)^2+(1-\alpha)(\gamma-1), \end{aligned}\[ which is a contradiction for large enough \(t\) and \(T\rightarrow \infty\). For the critical case \(\gamma=1+\frac{\beta}{d}\), we use the facts that \]u(t,x)^\gamma\geq u_1(t,x)^\gamma\[ and \]u(t,x)\geq u_2(t,x),\[ together with the estimates [\[Yvscampana\]](#Yvscampana){reference-type="eqref" reference="Yvscampana"} and [\[EstiPartLineal\]](#EstiPartLineal){reference-type="eqref" reference="EstiPartLineal"}. Therefore, for \(t>2\), we get \]\begin{aligned}
u(t,x)&\geq \int_1^{\frac{t}{2}} \int_{\mathbb{R}^d}Y(t-s,x-y)u(s,y)^{\gamma}dyds\\ &\geq C_1 C_2^\gamma\int_1^{\frac{t}{2}}(t-s)^{-\frac{\alpha d}{\beta}+\alpha-1}\int_{\mathbb{R}^d} e^{-\frac{\norm{x-y}^2}{4(t-s)}} s^{-\frac{\alpha d \gamma}{\beta}}e^{-\frac{\gamma\norm{y}^2}{s}}dyds\\ &=\frac{C_1 C_2^\gamma}{t^{\frac{d}{2}}}e^{-\frac{\norm{x}^2}{t}}\int_1^{\frac{t}{2}}\frac{(t-s)^{\alpha-1}t^{\frac{d}{2}}}{(t-s)^{\frac{d}{2}}s^{\frac{d+\beta}{2}}}\int_{\mathbb{R}^d} e^{\frac{\norm{x}^2}{t}-\frac{\norm{x-y}^2}{4(t-s)}-\frac{\gamma\norm{y}^2}{s}}dyds\\ &\geq\frac{C_1 C_2^\gamma}{t^{\frac{d}{2}}}e^{-\frac{\norm{x}^2}{t}}\int_1^{\frac{t}{2}}\frac{(t-s)^{\alpha-1}t^{\frac{d}{2}}}{(t-s)^{\frac{d}{2}}s^{\frac{d}{2}+\alpha}}\int_{\mathbb{R}^d} e^{\frac{\norm{x}^2}{t}-\frac{\norm{x}^2}{t}-\frac{\norm{y}^2}{2(t-s)}-\frac{\gamma\norm{y}^2}{s}}dyds\\ &\geq\frac{C_6}{t^{\frac{d}{2}}}e^{-\frac{\norm{x}^2}{t}}\int_1^{\frac{t}{2}}\frac{(t-s)^{\alpha-1}}{s^{\alpha}}ds\\ &\geq \frac{C_6}{t^{\frac{d}{2}}}e^{-\frac{\norm{x}^2}{t}}\int_1^{\frac{t}{2}}\frac{1}{t-s}ds \end{aligned}\[ and hence \]u(t,x)\geq \frac{C_6}{t^{\frac{d}{2}}}e^{-\frac{\norm{x}^2}{t}}\ln\left(2-\frac{2}{t}\right).\[ Using this and [\[F\]](#F){reference-type="eqref" reference="F"}, we obtain that \]\label{nuevacota} F(t)\geq \frac{C_7}{t^{\frac{d}{2}}}\ln\left(2-\frac{2}{t}\right).\[ Now, \]\begin{aligned}
t^{\frac{d}{2}\gamma}t^{1-\alpha} F(t)&=\frac{1}{2}t^{\frac{d}{2}\gamma}t^{1-\alpha} F(t)+\frac{1}{2}t^{\frac{d}{2}\gamma}t^{1-\alpha} F(t)\\ &\geq \frac{C_7}{2}\frac{t^{\frac{d}{2}\gamma}t^{1-\alpha}}{t^{\frac{d}{2}}}\ln\left(2-\frac{2}{t}\right)+\frac{C_5}{2}\int_0^t s^{\frac{d}{2}\gamma}F^\gamma(s) ds, \end{aligned}\[ where [\[nuevacota\]](#nuevacota){reference-type="eqref" reference="nuevacota"} yields the bound for the first term and the second term comes from [\[FacotadaCF\]](#FacotadaCF){reference-type="eqref" reference="FacotadaCF"}. The critical value of \(\gamma\) yields \]t^{\frac{d}{2}\gamma}t^{1-\alpha} F(t)\geq \frac{C_7}{2}t\ln\left(2-\frac{2}{t}\right)+\frac{C_5}{2}\int_0^t s^{\frac{d}{2}\gamma}F^\gamma(s) ds.\[ Defining the r.h.s. of this expression as the new \(f(t)\), \(t>1\), we proceed as before but using \]f(t)\geq C_8 t\ln\left(2-\frac{2}{t}\right)\[ and \]f'(t)\geq C_9 t^{\frac{d}{2}\gamma}F^\gamma(t)\[ instead of [\[cotaf\]](#cotaf){reference-type="eqref" reference="cotaf"} and [\[cotaf\'\]](#cotaf'){reference-type="eqref" reference="cotaf'"}, respectively, with \(C_8=\frac{C_7}{2}\) and \(C_9=\frac{C_5}{2}\). The resulting expression, instead of [\[exponentes\]](#exponentes){reference-type="eqref" reference="exponentes"}, is \]\frac{C_8^{1-\gamma}}{\gamma-1}t^{1-\gamma}\ln^{1-\gamma}\left(2-\frac{2}{t}\right)\geq C_9\int_t^T s^{-\frac{d}{2}\gamma(\gamma-1)-(1-\alpha)\gamma}ds\[ or, in this case, \]\frac{C_8^{1-\gamma}}{\gamma-1}t^{1-\gamma}\ln^{1-\gamma}\left(2-\frac{2}{t}\right)\geq C_9\int_t^T s^{-\gamma}ds.\[ This implies, as \(T\rightarrow\infty\), that \]C_8^{1-\gamma}\ln^{1-\gamma}\left(2-\frac{2}{t}\right)\geq C_9,\[ which is a contradiction whenever the initial condition is sufficiently large at the point \(y_0\).
The final result of this section deals with the case \(\gamma > 1+\frac{\beta}{d}\). For this purpose, as in Section [2](#sec:3){reference-type="ref" reference="sec:3"}, we set \]\kappa=\begin{cases}\frac{d}{\beta}, & d>\beta,\\ 1, & otherwise \end{cases}.\[ We also define \(H_2^\beta(\mathbb{R}^d):=\overline{C_0^\infty(\mathbb{R}^d)}^{\norm{\cdot}_{\Psi_{\beta},L_2}}\), with the clausure being respect to the graph norm \(\norm{\cdot}_{\Psi_{\beta},L_2}^2=\norm{\cdot}_{2}^2 + \norm{\Psi_{\beta}(-i\nabla)(\cdot)}_{2}^2\).
# Code availability {#code-availability .unnumbered}
Not applicable. | {'timestamp': '2022-02-07T02:05:29', 'yymm': '2109', 'arxiv_id': '2109.05305', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05305'} |
null | null |
# Brief Survey.
We survey the series convergence tests that have arisen in the literature. We begin with stating the classical ratio test found in standard calculus textbooks: **Theorem 1.1** (D'Alembert's Ratio Test) Let \(\{a_n\}\) be a sequence of positive numbers. Then
- If \(\limsup_{n\to \infty} \frac{a_{n+1}}{a_n} < 1\), then \(\sum_{n=1}^\infty a_n\) converges.
- If \(\liminf_{n\to \infty} \frac{a_{n+1}}{a_n} > 1\), then \(\sum_{n=1}^\infty a_n\) diverges.
- If \(\liminf_{n\to \infty} \frac{a_{n+1}}{a_n} \leq 1 \leq \limsup_{n\to \infty} \frac{a_{n+1}}{a_n}\), the test is inconclusive.
In conjunction with the root test, this test is often used for evaluation of the convergence of infinite series. Nevertheless, the test is not perfect since there exist series where the inconclusive case occurs, e.g. \(a_n = \frac{1}{n}\). As a result, several tests were developed to address this case such as Raabe's Test. **Theorem 1.2** (Raabe's Test) Let \(\{a_n\}\) be a sequence of positive numbers. Suppose we have \[\begin{aligned}
\frac{a_{n+1}}{a_n} = 1-\frac{\beta}{n} + \frac{\epsilon(n)}{n}, \end{aligned}\] where \(\beta\) is independent of \(n\) and \(\epsilon(n) \to 0\), as \(n\to \infty\). Then
- \(\sum_{n=1}^\infty a_n\) converges if \(\beta > 1\).
- \(\sum_{n=1}^\infty a_n\) diverges if \(\beta < 1\).
- If \(\beta = 1\), then test is inconclusive.
**Remark.** Notice that we may equivalently pose the test as: \(\sum_{n=1}^\infty a_n\) converges if \(\lim_{n\to \infty} n\left(1-\frac{a_{n+1}}{a_n}\right) > 1\) and \(\sum_{n=1}^\infty a_n\) diverges if \(\lim_{n\to \infty} n\left(1-\frac{a_{n+1}}{a_n}\right) < 1\). When \(\lim_{n\to \infty} n\left(1-\frac{a_{n+1}}{a_n}\right) = 1\), this corresponds to the inconclusive case. By weakening the assumption on the sequence term in Raabe's Test, one can prove a similar test due to Gauss. **Theorem 1.3** (Gauss' Test) Let \(\{a_n\}\) be a sequence of positive numbers. Suppose we have \[\begin{aligned}
\frac{a_{n+1}}{a_n} = 1-\frac{\beta}{n} + \frac{\gamma(n)}{n^{1+\lambda}}, \end{aligned}\] where \(\lambda > 0\), \(\beta\) is independent of \(n\) and \(\gamma(n)\) is bounded. Then
- \(\sum_{n=1}^\infty a_n\) converges if \(\beta > 1\).
- \(\sum_{n=1}^\infty a_n\) diverges if \(\beta \leq 1\).
We note that this test sufficiently determines convergence for any positive term series provided one can formulate the ratio term as in the expression given for Gauss' Test. There exists a test due to Kummer which contains as special cases the Ratio and Raabe's Test: **Theorem 1.4** (Kummer's Test) Let \(\{a_n\}\) be a sequence of positive numbers.
- Suppose there exists a sequence of positive numbers \(\{p_n\}\) and a positive number \(r\) such that \[\begin{aligned}
p_n\frac{a_n}{a_{n+1}}-p_{n+1} > r \end{aligned}\] holds for sufficiently large \(n\), then \(\sum_{n=1}^\infty a_n\) converges.
- On the other hand, if there exists a sequence of positive numbers \(\{p_n\}\) where \(\sum_{n=1}^\infty \frac{1}{p_n}\) diverges, such that \[\begin{aligned}
p_n\frac{a_n}{a_{n+1}}-p_{n+1} < r \end{aligned}\] for sufficiently large \(n\), then \(\sum_{n=1}^\infty a_n\) diverges.
By letting \(p_n = 1\) or \(n\), the test yields D'Alembert's Ratio Test or Raabe's Test, respectively. This test allows more flexibility in developing a series test by selecting appropriate sequences for \(p_n\). While the statement of the test is sufficient for convergence and divergence, Tong proves that the Kummer's Test in fact provides a characterization of convergence and divergence of series. Ali constructed the Second Ratio Test which can be used as a substitute for Raabe's and Gauss' Tests: **Theorem 1.5** (Second Ratio Test) Let \(\{a_n\}\) be a sequence of positive terms. Let \[\begin{aligned}
L = \max \{\limsup_{n\to \infty} \frac{a_{2n}}{a_n}, \limsup_{n\to \infty} \frac{a_{2n+1}}{a_n}\} \end{aligned}\] and \[\begin{aligned}
l = \min \{\liminf_{n\to \infty} \frac{a_{2n}}{a_n}, \liminf_{n\to \infty} \frac{a_{2n+1}}{a_n}\}. \end{aligned}\] Then
- If \(L < \frac{1}{2}\), then \(\sum_{n=1}^\infty a_n\) converges.
- If \(l > \frac{1}{2}\), then \(\sum_{n=1}^\infty a_n\) diverges.
- If \(l\leq \frac{1}{2} \leq L\), then the test is inconclusive.
Ali shows that the Second Ratio Test may be applied to a wide range of series that may require Raabe's or Gauss' Test. The advantage of this test is due to its computational efficiency for certain series in comparison to these tests. Next, we list results which are convenient for the proofs of Second Raabe Test and other associated tests. For monotonically decreasing sequences, we have the following result by Cauchy: **Theorem 1.6** (Cauchy's Condensation Test) Let \(\{a_n\}\) be a monotone sequence of positive numbers. Then \(\sum_{n=1}^\infty a_n\) converges if and only if \(\sum_{n=1}^\infty 2^na_{2^n}\) converges. Next, Ali proves a version of the Ratio Comparison Test using the two ratios in the Second Ratio Test: **Theorem 1.7** (Second Ratio Comparison Test) Let \(\{a_n\}\) and \(\{b_n\}\) be sequences of positive numbers. Suppose \[\begin{aligned}
\frac{a_{2n}}{a_n} \leq \frac{b_{2n}}{b_n} \quad \text{and} \quad \frac{a_{2n+1}}{a_n} \leq \frac{b_{2n+1}}{b_n} \end{aligned}\] for all large \(n\). Then
- \(\sum_{n=1}^\infty a_n\) converges if \(\sum_{n=1}^\infty b_n\) converges.
- \(\sum_{n=1}^\infty b_n\) diverges if \(\sum_{n=1}^\infty a_n\) diverges.
# Main Results.
**Theorem 2.1** (Second Raabe's Test) Suppose \(\{a_n\}\) is a sequence of positive numbers and \(\lambda \neq 0\). Let \[\begin{aligned}
\liminf_{n\to \infty} \ln{\lambda n}\left( \frac{1}{2}-\frac{a_{2n}}{a_n}\right) = m_1,\quad \liminf_{n\to \infty} \ln{\lambda n}\left( \frac{1}{2}-\frac{a_{2n+1}}{a_n}\right) = m_2,\\ \limsup_{n\to \infty} \ln{\lambda n}\left( \frac{1}{2}-\frac{a_{2n}}{a_n}\right) = M_1,\quad \limsup_{n\to \infty} \ln{\lambda n}\left( \frac{1}{2}-\frac{a_{2n+1}}{a_n}\right) = M_2. \end{aligned}\] Define \(m = \min \{m_1, m_2\}\) and \(M= \max \{M_1, M_2\}\).
- If \(m > \frac{\ln{2}}{2}\), then \(\sum_{n=1}^\infty a_n\) converges.
- If \(M < \frac{\ln{2}}{2}\), then \(\sum_{n=1}^\infty a_n\) diverges.
- If \(m\leq \frac{\ln{2}}{2} \leq M\), the test is inconclusive.
**Remark.** We note that the Second Raabe's Test involves \(\ln{n}\) as opposed to \(n\) in the Raabe's Test.
We will only prove ([\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) since the proof for ([\[eq2\]](#eq2){reference-type="ref" reference="eq2"}) will follow similarly. We rewrite ([\[eq1\]](#eq1){reference-type="ref" reference="eq1"}) and use L'Hopital's rule: \[\begin{aligned}
\lim_{n\to \infty} \ln{\lambda n}\left(\frac{1}{2}-\frac{(\ln{n})^p}{2(\ln{2n})^p}\right) &= \lim_{n\to \infty} \frac{1}{2}\left(\frac{1-\left(\frac{\ln{n}}{\ln{2n}}\right)^p}{\frac{1}{\ln{\lambda n}}}\right) \\ &= \frac{1}{2}\lim_{n\to \infty} \left( \frac{-p \left(\frac{\ln{n}}{\ln{2n}} \right)^{p-1} \left( \frac{\frac{\ln{2n}-\ln{n}}{n}}{(\ln{2n})^2}\right)}{-\frac{1}{(\ln{\lambda n})^2} \left( \frac{1}{n} \right)} \right) \\ &= \frac{1}{2}\lim_{n\to \infty} \left(p\left(\frac{\ln{n}}{\ln{2n}}\right)^{p-1}(\ln{2})\left(\frac{\ln{\lambda n}}{\ln{2n}}\right)^2 \right) \\ &= \frac{p\ln{2}}{2}. \end{aligned}\] The lemma has been proved. Suppose \(m\) and \(M\) are defined as in the statement of the theorem. We begin by proving *(i)*. Let \(m > \frac{\ln{2}}{2}\). WLOG, we assume \(m = m_1\). Therefore, we have \(m_1 \leq m_2\). Let \(\alpha, \beta \in \mathbb{R}\) such that \(\frac{\ln{2}}{2} < \beta < \alpha < m\). By assumption, there exists \(N_1 \geq 0\) such that \(\ln{\lambda n}\left( \frac{1}{2}-\frac{a_{2n}}{a_n}\right) > \alpha\), for all \(n\geq N_1\). Under algebraic rearrangement, we obtain \(\frac{a_{2n}}{a_n} < \frac{1}{2}-\frac{\alpha}{\ln{\lambda n}}\), for \(n \geq N_1\). Define \(b_n = \frac{1}{n\left(\ln{n} \right)^{\frac{2\beta}{\ln{2}}}}\). This allows us to obtain \(\frac{b_{2n}}{b_n} = \frac{(\ln{n})^{\frac{2\beta}{\ln{2}}}}{2(\ln{2n})^{\frac{2\beta}{\ln{2}}}}\). We invoke the lemma to obtain \[\begin{aligned}
\lim_{n\to \infty} \ln{\lambda n}\left(\frac{1}{2}-\frac{b_{2n}}{b_n} \right) &= \lim_{n\to \infty}\ln{\lambda n}\left(\frac{1}{2}-\frac{(\ln{n})^{\frac{2\beta}{\ln{2}}}}{2(\ln{2n})^{\frac{2\beta}{\ln{2}}}} \right) \\ &= \beta < \alpha. \end{aligned}\] Thus, there exists \(N_2 > 0\) such that \(\ln{\lambda n}\left(\frac{1}{2}-\frac{b_{2n}}{b_n} \right) < \alpha\) for all \(n \geq N_2\). Consequently, we have \(\frac{1}{2}-\frac{\alpha}{\ln{\lambda n}} < \frac{b_{2n}}{b_n}\). Let \(N_1^* = \max \{N_1, N_2\}\). Thus, for all \(n \geq N_1^*\), we obtain \(\frac{a_{2n}}{a_n} < \frac{1}{2}-\frac{\alpha}{\ln{\lambda n}} < \frac{b_{2n}}{b_n}\) which implies \(\frac{a_{2n}}{a_n} \leq \frac{b_{2n}}{b_n}\) for all \(n\geq N_1^*\). We shall next show that this same inequality holds for the corresponding ratios \(\frac{a_{2n+1}}{a_n}\) and \(\frac{b_{2n+1}}{b_n}\). Since \(m_2 > m_1 > \alpha\), by assumption there must exist \(N_3 \geq 0\) such that \(\ln{\lambda n}\left(\frac{1}{2}-\frac{a_{2n+1}}{a_n}\right) > \alpha\) for all \(n\geq N_3\). From here, we obtain \(\frac{a_{2n+1}}{a_n} < \frac{1}{2}-\frac{\alpha}{\ln{\lambda n}}\) for all \(n \geq N_3\). Using the same \(b_n\) as defined earlier, we get \(\frac{b_{2n+1}}{b_n} = \frac{n(\ln{n})^{\frac{2\beta}{\ln{2}}}}{(2n+1)(\ln{(2n+1)})^{\frac{2\beta}{\ln{2}}}}\). Using the lemma again, we have \[\begin{aligned}
\lim_{n\to \infty} \ln{\lambda n}\left( \frac{1}{2}-\frac{b_{2n+1}}{b_n}\right) = \beta < \alpha. \end{aligned}\] Thus, there exists \(N_4\) such that \(\ln{\lambda n}\left(\frac{1}{2}-\frac{b_{2n+1}}{b_n}\right) < \alpha\) for all \(n\geq N_4\). Hence, we obtain \(\frac{1}{2}-\frac{\alpha}{\ln{\lambda n}} < \frac{b_{2n+1}}{b_n}\) for all \(n\geq N_4\). If we let \(N_2^* = \max \{N_3, N_4\}\), we have \(\frac{a_{2n+1}}{a_n} \leq \frac{b_{2n+1}}{b_n}\) for all \(n\geq N_2^*\). Finally, if we let \(N = \max \{N_1^*, N_2^*\}\), we have that both \(\frac{a_{2n}}{a_n} \leq \frac{b_{2n}}{b_n}\) and \(\frac{a_{2n+1}}{a_n} \leq \frac{b_{2n+1}}{b_n}\) hold for \(n \geq N\). Notice that \(\sum_{n=2}^\infty b_n\) converges since \(\sum_{n=2}^\infty \frac{1}{n(\ln{n})^p}\) converges for any \(p > 1\). Thus, by the Second Ratio Comparison Test, \(\sum_{n=1}^\infty a_n\) must converge. Hence, we have proved *(i)*. We shall now prove *(ii)*. The proof will more or less imitate the proof for *(i)*. Suppose \(M < \frac{\ln{2}}{2}\). WLOG, we assume \(M = M_1\). Therefore, we have \(M_2 \leq M_1\). Let \(\alpha, \beta \in \mathbb{R}\) such that \(M < \alpha < \beta < \frac{\ln{2}}{2}\). By assumption, there exists \(N_1 \geq 0\) such that \(\ln{\lambda n}\left( \frac{1}{2}-\frac{a_{2n}}{a_n}\right) < \alpha\), for all \(n\geq N_1\). Under algebraic rearrangement, we obtain \(\frac{1}{2}-\frac{\alpha}{\ln{\lambda n}}< \frac{a_{2n}}{a_n}\), for \(n \geq N_1\). Define \(b_n = \frac{1}{n\left(\ln{n} \right)^{\frac{2\beta}{\ln{2}}}}\). This allows us to obtain \(\frac{b_{2n}}{b_n} = \frac{(\ln{n})^{\frac{2\beta}{\ln{2}}}}{2(\ln{2n})^{\frac{2\beta}{\ln{2}}}}\). We invoke the lemma to obtain \[\begin{aligned}
\lim_{n\to \infty} \ln{\lambda n}\left(\frac{1}{2}-\frac{b_{2n}}{b_n} \right) &= \lim_{n\to \infty}\ln{\lambda n}\left(\frac{1}{2}-\frac{(\ln{n})^{\frac{2\beta}{\ln{2}}}}{2(\ln{2n})^{\frac{2\beta}{\ln{2}}}} \right) \\ &= \beta > \alpha. \end{aligned}\] Thus, there exists \(N_2 > 0\) such that \(\ln{\lambda n}\left(\frac{1}{2}-\frac{b_{2n}}{b_n} \right) > \alpha\) for all \(n \geq N_2\). Consequently, we have \(\frac{b_{2n}}{b_n} < \frac{1}{2}-\frac{\alpha}{\ln{\lambda n}}\). Let \(N_1^* = \max \{N_1, N_2\}\). Thus, for all \(n \geq N_1^*\), we obtain \(\frac{b_{2n}}{b_n} < \frac{1}{2}-\frac{\alpha}{\ln{\lambda n}} < \frac{a_{2n}}{a_n}\) which implies \(\frac{b_{2n}}{b_n} \leq \frac{a_{2n}}{a_n}\) for all \(n\geq N_1^*\). We shall next show that this same inequality holds for the corresponding ratios \(\frac{a_{2n+1}}{a_n}\) and \(\frac{b_{2n+1}}{b_n}\). Since \(M_2 \leq M_1 < \alpha\), by assumption there must exist \(N_3 \geq 0\) such that \(\ln{\lambda n}\left(\frac{1}{2}-\frac{a_{2n+1}}{a_n}\right) < \alpha\) for all \(n\geq N_3\). From here, we obtain \(\frac{a_{2n+1}}{a_n} > \frac{1}{2}-\frac{\alpha}{\ln{\lambda n}}\) for all \(n \geq N_3\). Defining \(b_n\) as earlier, we get \(\frac{b_{2n+1}}{b_n} = \frac{n(\ln{n})^{\frac{2\beta}{\ln{2}}}}{(2n+1)(\ln{(2n+1)})^{\frac{2\beta}{\ln{2}}}}\). Using the lemma again, we have \[\begin{aligned}
\lim_{n\to \infty} \ln{\lambda n}\left( \frac{1}{2}-\frac{b_{2n+1}}{b_n}\right) = \beta > \alpha. \end{aligned}\] Thus, there exists \(N_4\) such that \(\ln{\lambda n}\left(\frac{1}{2}-\frac{b_{2n+1}}{b_n}\right) > \alpha\) for all \(n\geq N_4\). Hence, we obtain \(\frac{b_{2n+1}}{b_n}<\frac{1}{2}-\frac{\alpha}{\ln{\lambda n}}\) for all \(n\geq N_4\). If we let \(N_2^* = \max \{N_3, N_4\}\), we have \(\frac{b_{2n+1}}{b_n} \leq \frac{a_{2n+1}}{a_n}\) for all \(n\geq N_2^*\). Finally, if we let \(N = \max \{N_1^*, N_2^*\}\), we have that both \(\frac{b_{2n}}{b_n} \leq \frac{a_{2n}}{a_n}\) and \(\frac{b_{2n+1}}{b_n} \leq \frac{a_{2n+1}}{a_n}\) hold for \(n \geq N\). Notice that \(\sum_{n=2}^\infty b_n\) diverges since \(\sum_{n=2}^\infty \frac{1}{n(\ln{n})^p}\) diverges for any \(p < 1\). Thus, by the Second Ratio Comparison Test, \(\sum_{n=1}^\infty a_n\) diverges. Hence, we have proved *(ii)*. ◻
**Example 1. (The harmonic series)** Consider \(a_n = \frac{1}{n}\). Notice that \[\begin{aligned}
\frac{a_{2n}}{a_{n}} = \frac{1}{2} \to \frac{1}{2},\quad \text{as \(n\to \infty\)} \end{aligned}\] and \[\begin{aligned}
\frac{a_{2n+1}}{a_{n}} = \frac{n}{2n+1} \to \frac{1}{2},\quad \text{as \(n\to \infty\)}. \end{aligned}\] Thus, the Second Ratio Test fails to give a conclusion. We will use the Second Raabe's Test: \[\begin{aligned}
\lim_{n\to \infty}\ln{n}\left( \frac{1}{2}-\frac{1}{2}\right) = 0 \end{aligned}\] and \[\begin{aligned}
\lim_{n\to \infty} \ln{n}\left( \frac{1}{2}-\frac{n}{2n+1}\right) = 0. \end{aligned}\] Since \(M = 0\), we conclude by the Second Raabe's Test that \(\sum a_n\) diverges. Alternatively, we may use Corollary 2.1.2. where \(\beta = 0\) to conclude \(\sum a_n\) diverges. We now state and prove a generalized Kummer's Test that uses the quantities involved with the Second Ratio Test: **Theorem 2.2** (Second Kummer's Test) Let \(\{a_n\}\) be a sequence of positive numbers.
- Suppose there exists a sequence of positive numbers \(\{p_n\}\) and positive numbers \(r_1\) and \(r_2\) such that \[\begin{aligned}
p_n\frac{a_n}{a_{2n}}-2p_{2n} > r_1 \end{aligned}\] and \[\begin{aligned}
p_n\frac{a_n}{a_{2n+1}}-2p_{2n+1} > r_2 \end{aligned}\] where both inequalities hold for sufficiently large \(n\), then \(\sum_{n=1}^\infty a_n\) converges.
- On the other hand, if there exists a sequence of positive numbers \(\{p_n\}\) where \(\sum_{k=1}^\infty \frac{1}{2^kp_{2^k}}\) diverges, and at least one of the following inequalities holds for sufficiently large \(n\) \[\label{eq3} p_n\frac{a_n}{a_{2n}}-2p_{2n} < 0\] or \[\label{eq4} p_n\frac{a_n}{a_{2n+1}}-2p_{2n+1} < 0\] then \(\sum_{n=1}^\infty a_n\) diverges.
**Remark.** If *(i)* holds for the sequence \(p_n = 1\), we obtain the following corollary. **Corollary 2.1.3** Suppose \(\{a_n\}\) is a sequence of positive numbers. If the following inequalities hold for positive numbers \(r_1, r_2\) and \(n\) sufficiently large: \[\begin{aligned}
\frac{a_{2n}}{a_n} < \frac{1}{2+r_1}\quad \text{and}\quad \frac{a_{2n+1}}{a_n} < \frac{1}{2+r_2}, \end{aligned}\] then \(\sum_{n=1}^\infty a_n\) converges. Thus, we obtain a stricter form of the Second Ratio Test, since if the above conditions are met, then \(L = \max \{ \limsup \frac{a_{2n}}{a_n}, \limsup \frac{a_{2n+1}}{a_n}\} < \frac{1}{2}\). On the other hand, letting \(p_n = \frac{1}{n}\) in *(ii)*, we obtain another corollary. **Corollary 2.1.4** Suppose \(\{a_n\}\) is a sequence of positive numbers. If the either one of the following inequalities hold for \(n\) sufficiently large: \[\begin{aligned}
\frac{a_{2n}}{a_n} > 1\quad \text{or}\quad \frac{a_{2n+1}}{a_n} > 1 + \frac{1}{2n}, \end{aligned}\] then \(\sum_{n=1}^\infty a_n\) diverges. In comparison to the Second Ratio Test, it is sufficient to check whether one of the two ratios satisfies its corresponding inequality. In this way, this test is more general compared to the Second Ratio Test for divergence.
# Conclusion
The novelty of the Second Raabe's Test is that it addresses the inconclusive case found in the Second Ratio Test; nevertheless, like the Raabe's Test, there is also an inconclusive case. In this case, several tests were developed (the Second Gauss' Test, Second Kummer's Test) to partially address this case. However, more work could be done to further generalize the Second Gauss's Test and Second Kummer's Test, as we suspect that there exists a more general formulation. In particular, we speculate that there exists a characterization of convergence and divergence of series based on the Second Kummer's Test, as was shown in the ordinary Kummer's Test. Ali constructed the \(m\)-th ratio test, which generalizes the Second Ratio Test from using two ratios to \(m\) ratios, to test series convergence. While perhaps not too hard to formulate, another potential research direction would be the construction of series tests that address the inconclusive case found in the \(m\)-th ratio test. **Acknowledgment.** The author wishes to thank Professor Zhonghai Ding for his guidance on this work. The author would also like to thank Keoni Castellano, Michael Schwob, Shen Huang, and Bowen Liu for reading and proofreading the manuscript. | {'timestamp': '2021-09-16T02:05:59', 'yymm': '2109', 'arxiv_id': '2109.05447', 'language': 'en', 'url': 'https://arxiv.org/abs/2109.05447'} |
Subsets and Splits